Loading Now
Artificial Intelligence Computation and Language Computer Vision , , , , , ,

Fine-Tuning Frontiers: Unleashing AI’s Potential Across Domains

Latest 100 papers on fine-tuning: Jul. 18, 2026

The landscape of AI/ML is constantly evolving, with fine-tuning playing a pivotal role in adapting powerful foundation models to specific, often challenging, tasks. Recent research showcases remarkable breakthroughs, demonstrating how targeted fine-tuning strategies, coupled with novel architectural designs and innovative data approaches, are pushing the boundaries of what AI can achieve. From making large language models safer and more efficient to enabling robots to perform complex real-world tasks, these advancements highlight a significant shift towards more adaptable, robust, and domain-aware AI systems.

The Big Ideas & Core Innovations

At the heart of these advancements lies the ingenuity in solving core problems. One prevalent theme is tackling the limitations of fixed, monolithic models. For instance, in “Multi-Head Latent Control: A Unified Interface for LLM Agent Decision Making” by Amirhosein Ghasemabadi and their colleagues at the ECE Department, University of Alberta, and Huawei Technologies Canada Co., Ltd., a lightweight layer is introduced to enable frozen LLMs to make deployment-time decisions, significantly reducing API costs while improving performance. This is echoed in “SymbOmni: Evolving Agentic Omni Models via Symbolic Concept Learning” by the South China University of Technology, Westlake University, Johns Hopkins University, The Chinese University of Hong Kong, Shenzhen Loop Area Institute, and SphereLab, which addresses the ‘perpetual novice’ problem by enabling visual generation models to learn cumulatively through reusable symbolic workflow instructions, bypassing gradient-based fine-tuning for continuous self-improvement.

Another critical area is enhancing model robustness and safety, particularly when fine-tuning. “DataShield: Uncovering Risky Fine-Tuning Data Across LLMs Through Consensus Subspace Alignment” from Zhejiang University, UESTC, and Shanghai AI Laboratory introduces a data-centric framework to identify and mitigate risky fine-tuning samples, drastically reducing attack success rates. Similarly, “PersGuard: Preventing Malicious Personalization in Text-to-Image Diffusion Models via Model Backdoors” by Xinwei Liu and colleagues from the Chinese Academy of Sciences and Nanyang Technological University proposes a backdoor-based protection to safeguard text-to-image diffusion models from unauthorized personalization, ensuring defensive behaviors are triggered upon misuse.

Efficiency and resource optimization are also central. The paper “Long-Context Fine-Tuning with Limited VRAM” by Vladimir Fedosov and the BMW Group combines Hierarchical Global Attention with segment-wise backpropagation and tiered KV storage to enable long-context fine-tuning of large language models on GPUs with limited VRAM. For low-resource scenarios, “DeltaMerge-LowRes: Composing Language and Task Deltas for Low-Resource Adaptation” by Son Ha Xuan and colleagues from RMIT University and Ho Chi Minh City University of Technology introduces cross-axis TIES merging to combine language and task deltas, significantly improving performance on summarization and QA in African languages without joint fine-tuning.

Specialized applications are seeing tailored fine-tuning strategies. In robotics, “Semantic Anchoring for Robotic Action Representations” by Yuan Xu, Youheng Shi, Chengyang Li, Wentao Zhu, and Yizhou Wang from Peking University and Eastern Institute of Technology, Ningbo, proposes a plug-and-play method to preserve the semantic structure of action representations, improving out-of-distribution generalization in bimanual robots. For medical imaging, “Demographically-Conditioned Synthetic Medical Images for Bias Mitigation and Bias Detection in Disease Classifiers” by Mahmoud Ibrahim and collaborators at Maastricht University and VITO, Belgium, uses fine-tuned Stable Diffusion to generate synthetic data, both mitigating and detecting bias in medical image classifiers with remarkable data efficiency. “scVision: A Vision Foundation Model for Single-Cell Biology via Spatial Gene Cartography” by Ridvan Yesiloglu, Sakib Mostafa, James Zou, Ash Alizadeh, Jiajun Wu, Lei Xing, Ehsan Adeli, and Md Tauhidul Islam from Stanford University, transforms single-cell transcriptomics into images, allowing a vision foundation model to achieve state-of-the-art zero-shot cell-type annotation and gene-program discovery.

Under the Hood: Models, Datasets, & Benchmarks

These innovations are often built upon advancements in core components and rigorous evaluation. Here are some key resources and methodologies:

  • VEXMLM: A vocabulary-extended XLM-RoBERTa for Ge’ez-script languages, utilizing language-specific SentencePiece tokenizers and mean-based embedding initialization. Code: https://github.com/hailaykidu/VEXMLM.
  • HGA (Hierarchical Global Attention): Employed with Qwen3-8B and PG19 for long-context fine-tuning with limited VRAM. Code: https://github.com/vfedosov77/HierarchicalGlobalAttention.
  • DataShield: Constructs consensus safe/unsafe subspaces using semantic spectral decomposition on multiple safety-aligned LLMs. Code: https://github.com/ZJU-LLM-Safety/DataShield.
  • Digital Pantheon: Uses Gemma3 (27B) and Qwen3.6 (27B) models, with SFT+DPO+RAG, grounded in 2019 Flemish party manifestos for political simulations.
  • CosFly-VLA: A vision-language-action model for UAV tracking, built on a frozen Qwen3.5 backbone and trained with spatially grounded continued pretraining and RL.
  • SD-MAR: Synthetic data generation framework for multi-image analytical reasoning, training Vision Language Models (Qwen2.5-VL-7B) with GRPO-lite.
  • CFM-Bench: A unified benchmark for Channel Foundation Models, encompassing six data domains (statistical simulation, ray-tracing, measured terrestrial/aerial channels, multimodal vehicular) and six task groups (PHY, RAN, ISAC).
  • SFF-CLIP: Fine-tunes CLIP using self-annotated region alignment based on text-specific heat maps, improving fine-grained visual understanding.
  • WrAFT: A modular LLM-powered (GPT-4o, Claude 3.7) automated writing evaluation system for argumentative essays, achieving SOTA scoring and layered feedback. Code: https://github.com/judywq/wraft.
  • BT5: A T5-small model, fine-tuned on the NIKL Korean Print-Braille Parallel Corpus, outperforming large LLMs for Braille translation.
  • Multi-Head Latent Control: Lightweight layer attaching to frozen LLMs/VLMs like Qwen3-VL-8B. Code: https://github.com/Amirhosein-gh98/Multi-Head-Latent-Control.
  • scVision: A vision foundation model for single-cell biology, pre-trained on 72 million human cells with masked image modeling.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • Exploration with Guided Search (EXPLORE): Combines simulator-guided MCTS with transformer-based decoding for analog circuit topology generation.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • CSSEL-P2P: A data-driven approach for simultaneous speech translation with decoder-only LLMs using fixed-length chunks and prefix-to-prefix targets.
  • CHUNKLOW: A seam-aware training framework for chunked policies in VLA robotic models, achieving continuity-consistent action transitions on CALVIN and LIBERO.
  • REAL: A closed-loop framework for training tool-calling embodied agents, using SFT and GSPO RL on Qwen3-VL-8B, with sim-to-real transfer to Ark LIFT2.
  • OvisOCR2: A compact 0.8B end-to-end document parser, trained with SFT, RL, and on-policy distillation from a 4B teacher (Qwen3.5 backbone). Code: https://huggingface.co/ATH-MaaS/OvisOCR2.
  • TRACE: A critic-free credit-assignment method for long-horizon agentic RL, using a frozen reference model (Qwen3-4B, Qwen3-30B-A3B) for turn-level rewards.
  • CARE-LoRA: Memory-efficient LoRA fine-tuning for LLMs, NLU, and diffusion models, with compressed activation reconstruction. Code: https://github.com/fishandyu/CARE-LoRA.
  • FeDiSyn: A framework for federated fine-tuning of large vision models, using Stable Diffusion v1-5 and LLaMa-7B for synthetic pre-training.
  • DermDepth: The first single-view metric scale 3D reconstruction model for dermatology, trained on the D-Synth synthetic dermoscopic dataset. Code: https://github.com/hectorcarrion/dermdepth.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: https://github.com/mcxiaoxiao/QDA-SQL.
  • TALKE: A text dataset distillation framework using trajectory-integrated influence functions and Optimal Transport. Code: https://github.com/votrinhan88/take.
  • Unified Multi-Domain OCR (Manchu): Reuses fine-tuning checkpoints as domain specialists for historical Manchu OCR, routing with a ResNet-18 image classifier.
  • TA-RS: Traffic-Aware Randomized Smoothing for LLM-based network intrusion detection, achieving certified robustness on CIC-IDS-2018 and HIKARI-2021.
  • CARE-PPO: A PPO-based RL framework for language-based quantitative prediction with confidence estimation, evaluated on Qwen-3 models.
  • LLPR (Last-Layer-Projection Regression): Uncertainty quantification for MLFFs, integrated into foundation model fine-tuning with MACE-MP-0. Code: https://github.com/chenggroup/ai2-kit.
  • EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models.
  • QDA-SQL: A data augmentation method for multi-turn Text-to-SQL tasks, generating QA pairs using LLMs, integrated with a StateFlow framework. Code: [https://github.com/mcxiaox

Share this content:

mailbox@3x Fine-Tuning Frontiers: Unleashing AI's Potential Across Domains
Hi there 👋

Get a roundup of the latest AI paper digests in a quick, clean weekly email.

Spread the love

Discover more from SciPapermill

Subscribe to get the latest posts sent to your email.

Tag

Post Comment

SciPapermill: Follow the latest research. Copyright 2026 | Powered By SpiceThemes

Discover more from SciPapermill

Subscribe now to keep reading and get access to the full archive.

Continue reading