From Depth to Dialog: Unpacking the Latest in Foundation Models

Latest 100 papers on foundation models: Oct. 17, 2026

Foundation models are reshaping the AI landscape, demonstrating unprecedented capabilities across diverse domains. From revolutionizing how we understand and interact with the physical world to enabling hyper-efficient data analysis and accelerating scientific discovery, these models are constantly pushing the boundaries of what’s possible. This digest dives into recent breakthroughs, highlighting core innovations and their profound implications.

The Big Idea(s) & Core Innovations

Recent research underscores a dual push: making foundation models more adaptable and efficient for specialized tasks, and enhancing their capacity for sophisticated reasoning and interaction.

One major theme is efficient adaptation and specialized grounding. For instance, ControlFlow3D: Distilling Multi-View Knowledge into Latent Flow Matching for Point Cloud Upsampling from Nanjing University of Aeronautics and Astronautics proposes distilling knowledge from multi-view image models into lightweight 3D generation networks, enabling efficient point cloud upsampling without multi-view images at inference. Similarly, DeFM: Learning Foundation Representations from Depth for Robotics by researchers from ETH Zurich and Stanford introduces a self-supervised foundation model trained exclusively on depth images, proving that robust semantic understanding can emerge even without color or texture, crucial for robotics. In medical imaging, the GDBA-Net: A Gated Dual-Branch Adaptation of SAM2 for Robust Cross-Domain Nuclei Segmentation by the University of Missouri and colleagues, efficiently adapts SAM2 for robust nuclei segmentation across diverse datasets by combining semantic and frequency-derived structural priors with a sample-adaptive gating mechanism. Complementing this, FAD-ViT: Learning Features, Not Layouts: Feature-Axis Attention as a Transfer Prior for Medical Imaging from University College Dublin shows that decoupling spatial and feature attention, with feature-axis attention as the primary global operator, drastically improves medical image transfer learning from natural image pretraining.

Another critical area is enhancing reasoning and agentic capabilities. The ARC: A Reasoning Recipe for Robot Foundation Models by the University of Illinois Urbana-Champaign and NVIDIA significantly improves zero-shot robot task performance by introducing action-grounded causal reasoning traces, making RFMs robust to vague instructions. Similarly, PhySGA: A Physics-Space Grounded Agentic Framework for Generalizable and Interactive 3D Composition from Tsinghua University leverages agent exploration and physical simulation for universal interactive 3D object assembly, ensuring strict physical validity. In a fascinating development for multi-agent systems, Mental-Models for Multi-Agent Systems from the University of California, San Diego, proposes agents that learn amortized recursive ‘Theory-of-Mind’ representations to infer other agents’ beliefs and intentions, improving interaction quality. The challenge of handling harmful or unreliable knowledge in continuous learning is tackled by Reliable Knowledge Acquisition for Continual Learning Agents (RIKA) by the University of California, Davis, and University of Illinois Urbana-Champaign, which uses evidence-based admission for knowledge reuse.

For time-series and tabular data, the focus is on robustness and efficiency. PathFM: Fast Generative Forecasting of Multivariate Trajectories from Caltech and Amazon introduces a flow-matching model for fast and calibrated trajectory generation, capturing joint temporal and cross-variable dependencies for superior path quality. SurvICL: a Foundation Model for in-context Survival Analysis by Inria de Saclay and colleagues, sets a new benchmark in survival analysis by outperforming the classical Cox model on small datasets using a proportional-odds prior and quantile regression. In power systems, Learning by Examples: Power Grid Steady State Analysis with a Tabular Foundation Model by Delft University of Technology shows that pretrained tabular foundation models (TabPFN) can act as general-purpose neural solvers for power flow and contingency screening without retraining. However, tabular models also face unique challenges, with Tabular Foundation Models Are Effectively Shallow from ETH Zurich demonstrating that TFMs are vastly overparameterized, with a simple linear map able to replace up to 96% of transformer blocks with minimal performance loss, leading to significant speedups and memory reductions.

Specialized applications and novel architectures continue to emerge. LWM-TF: A Wireless Foundation Model on 5G NR Time-Frequency Resource Grids by Arizona State University, pre-trains a factorized masked autoencoder on 5G NR grids, enabling label-efficient transfer learning for tasks like modulation classification and mobility detection. For computational biology, Scaling RNA Foundation Models through Iterative Refinement (Eco-RNA) by The Chinese University of Hong Kong introduces an 80M-parameter RNA foundation model that leverages iterative refinement to surpass much larger models, showing that internal test-time computation can advance the compute-performance frontier. Even the very act of creating foundation models is getting an upgrade with MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal Data from the University of California, Riverside, which uses proxy textual descriptions as semantic bridges to enable multimodal LLMs to reason across modalities without real paired cross-modal data.

Under the Hood: Models, Datasets, & Benchmarks

Innovations in foundation models often come hand-in-hand with new datasets, benchmarks, and architectural insights. Here’s a quick look at key resources and techniques:

• PathFM: Utilizes flow matching with Parallel Decoding Distillation (PDD) for fast trajectory generation. Introduces the Forecast Path Score Bench (fps-bench) for path distribution evaluation, alongside existing fev-bench and TIME benchmarks. Code and model weights are planned for release.
• GeoState: Learns view-decoupled geometry states using a Scene Tokenizer and Geometry Decoder, applied to datasets like Waymo Open Dataset, KITTI, and nuScenes. Project page: https://hongchang-liu.github.io/GeoState-project-page/.
• SurvICL: A tabular foundation model using a novel proportional-odds prior and plain quantile regression. Evaluated on the largest survival benchmark to date with 103 datasets. Code and model checkpoints to be open-sourced.
• Pathology foundation models (Atlas2, Atlas2-B, Atlas2-S, UNI2h): Used in a training-free content-based retrieval system for whole slide images, leveraging a modified Chamfer similarity metric. Retrieval pipeline code available upon request.
• SelectPFN: A two-pass Prior-Data Fitted Network (PFN) for amortized Bayesian variable selection. Achieves exact finite-sample FDR control with model-X knockoffs. No public code specified.
• Grid Foundation Model (GridFM): Distilled into a smaller student model using physics-aware distillation. A renewable-grid benchmark based on IEEE 118-bus system is introduced. Uses PyTorch Geometric and references PGLib-OPF and gridfm-datakit-v1.
• CXR Foundation Models: Benchmarked for adult-to-pediatric domain shift, including RAD-DINO, MedCLIP, BiomedCLIP. Evaluated on VinDr-CXR (adult) and PediCXR (pediatric). Code to be publicly available.
• RIKA (Reliable Knowledge Acquisition): Combines hierarchical knowledge extraction with Evidence-based Knowledge Admission (EKA). Evaluated across multiple LLM backbones (GPT-5.4, DeepSeek-V4-Flash, MiniMax-M3, Gemini-3-Flash) on Continual Learning Bench. Code not specified.
• Protein Language Models (ProtGPT2, ProGen2): Attacked via pre-training and retrieval corpus poisoning using UniProt/UniRef databases. No public code specified.
• ControlFlow3D: Distills multi-view knowledge into a latent flow matching network with a Mamba-based geometry decoder. Introduces ShapeNetPU, a large-scale multi-modal benchmark. Code available: https://github.com/I2-Multimedia-Lab/ControlFlow3D.
• LWM-TF: A factorized masked autoencoder pre-trained on 5G NR time-frequency grids. Uses a 3GPP-compliant 1M-slot PUSCH dataset. No public code specified.
• PiCo: Post-hoc explainability for Geo-Foundation Models (GFMs) using synthetic multispectral data from CAMP3D with 3D radiative transfer. Code available: https://github.com/yihshe/PiCo.git.
• MSP: Trains multimodal LLMs using proxy textual descriptions as semantic bridges. Leverages encoders like LAION-CLAP and MetaCLIP ViT-B/16, with Qwen3-8B as the language model. Benchmarked on AVQA, MUSIC-AVQA, OmniBench. No public code specified.
• TEAL: Active learning for semantic segmentation using DINOv3 representations and a Mini-DPT segmentation head. Evaluated on CamVid, Cityscapes, ADE20K, Pascal Context. No public code specified.
• EV-JEPA: An elastic Vision-JEPA model with Transition-Aware Triplet Training (T3) for cross-device inference. Built on V-JEPA 2 and evaluated on SSv2, Kinetics-710, HowTo100M, Epic-Kitchens 100. No public code specified.
• PhySGA: Agentic framework for 3D composition with dual-agent collaboration and a Self-Evolving Skill Library. Uses Isaac Sim. Project page: https://hongchang-liu.github.io/GeoState-project-page/.
• TabPFN: Applied to power grid steady-state analysis, with TabPFN-R and TabPFN-FD variants. Uses PFΔ benchmark dataset and GridFM datakit. Code: TabPFNv3, GridFM datakit (gridfm-datakit-v1).
• Atomic-WAM: Connects task reasoning to video-action generation using atomic-motion descriptions and an Execution-Grounded Reasoner Adaptation (EG-GRPO). Uses DROID dataset and RoboLab benchmark. Project page: https://ssj9596.github.io/Atomic-WAM-project.
• SMORE (Scaffold MemOry for Revision): A revisable external memory for long-horizon 3D geometry, model-agnostic and compatible with existing GFMs like DUSt3R, MASt3R, VGGT. Code, data, and models will be publicly available.
• Test-Time Forecasting (TTF): Leverages recurrent depth and theoretical links to Koopman operator theory for in-context dynamics prediction. Introduces implicit TTF (iTTF) and explicit TTF (eTTF). No public resources or code specified.
• Spectral World Renderer (SWR): Unifies 3D reconstruction and hyperspectral unmixing using Gaussian splatting. Introduces SWR-Bench with 8 synthetic worlds. Project page: https://ming053l.github.io/Spectral-World-Renderer/.
• Kandinsky S2V 1.0: Audio-driven avatar video generation with 2B Lite and 21B Pro models. Uses Qwen2.5-Omni for conditioning and FramePack for visual memory. Benchmarked on EMTD, CelebV-HQ, and HDTF. Code and model weights: https://github.com/kandinskylab/kandinsky-s2v.
• λ0: A humanoid foundation model for loco-manipulation trained on HUMANVERSE, a 500-hour dataset of egocentric human video with synchronized motion. Evaluated on SIMPLE benchmark tasks and Unitree G1 robot. Code: https://github.com/LambdaHumanoid/openlambda.
• RT-J: An 85M-parameter relational foundation model trained on THE JOIN, a corpus of 639 web-crawled databases. Uses random-walk retrieval. Code: https://github.com/stanford-star/agentic-task-gen and project page: https://star-project.stanford.edu/rt-j.
• ANCHOR: A benchmark for evaluating engineering constraint following in coding agents. Uses Claude Opus 4.8 and assesses compliance across a 10-category taxonomy. Dataset: https://huggingface.co/datasets/Yone-01/ANCHOR. Code: https://github.com/bupterlxp/ANCHOR.
• SuperAI: A cross-architecture multi-model fusion framework for integrating capabilities into a single host model (e.g., Llama-3.2-3B, Qwen2.5-3B, Pythia-2.8B). No public code specified.
• TTCO (Test-Time Command Optimization): Adapts commands to pretrained humanoid Behavior Foundation Models (BFMs) to reduce execution errors. Code and example data provided as supplementary material.
• EMG-FM-BENCH: A benchmark for PPG foundation models on EMG tasks, unifying 20 public datasets. Evaluates models like MOMENT, Chronos-2, TimesFM, BRANT, Lag-Llama, UniTS, PatchTST. Code: https://anonymous.4open.science/r/EMG-FoundBench/.
• RRM-GPT: An autoregressive framework for radio resource management in 5G NR networks, applying language model principles. No public resources or code specified.
• Colon3R: Cross-domain semi-supervised framework for monocular colonoscopic 3D reconstruction, adapting VGGT foundation model. Uses C3VD and SimCol3D benchmarks. Code: https://github.com/vglx/Colon3R.
• TabFORGE: A tabular foundation model for generative modeling using a PFN-based encoder, score-based diffusion transformer, and denoising-aligned decoder. Achieves SOTA on 45 datasets. Code: https://github.com/SilenceX12138/TabFORGE. Hugging Face: https://huggingface.co/XiangjianJiang/TabFORGE.
• TABMF: Repurposes pretrained TFMs like TabICLv2 and TabPFN-2.5 as encoders of supervised datasets, using contrastive objectives on synthetic views. Evaluated on retrieval and distribution-shift tasks. No public code specified.
• Geospatial Foundation Models (TerraMind, Prithvi-EO-2.0, DOFA, DINOv3): Investigated for shortcut learning in cross-disaster damage assessment on xView2 and BRIGHT datasets. Code: https://github.com/alishibli97/gfm-shortcut-learning.
• Calendar Encoding for TSFMs (Chronos-2, TabPFN-TS-3, TiRex-2, TimesFM-3): Systematic comparison of 7 encoding strategies on 9 diverse datasets. Code to be released.
• TabReliBench: Benchmark for tabular foundation models on reliability metrics beyond accuracy. Uses 160 classification and 81 regression datasets from OpenML. No public code specified.
• PPG-Bench: Unified benchmark for PPG (photoplethysmography) foundation models (AnyPPG, PaPaGei-S/P, PulsePPG, SigmaPPG) across 9 datasets and 26 tasks. Code: https://github.com/mannas-ai/ppg-bench.
• PlacePulse: Geospatial foundation model learning from structural attributes (demographic, economic) using H3 neighborhood aggregates. Compared with Google Population Dynamics Foundation Model (PDFM). No public code specified.
• Chronos-Bolt and TimesFM 2.5: Investigated for structured block skipping to reduce inference costs in time-series forecasting. No public code specified.
• Eco-RNA: 80M-parameter RNA foundation model with iterative refinement. Evaluated on RNAGym benchmark. Model weights and code to be released.
• A4-MR: 3D vision-language foundation model for MRI with abnormality-aware contrastive loss. Based on DINOv2, DETR, CLIP. Uses datasets like CirrMRI600+, AMOS22. No public code specified.
• Brain MRI Foundation Models (3DINO, SAM-Med3D, BrainIAC, BrainFM): Diagnosed for cross-dataset failure using 17 public brain MRI datasets. Code: https://github.com/BrainFM/brainfm-profiler/tree/main/representation_utility.
• Diffusion Meta-Prompting (DMP): Uses diffusion models to generate task-specific prompts for foundation models. Works with CLIP, Stable Diffusion. No public code specified.
• Fluid-Gen-Zero: Training-free framework injecting physics awareness into pretrained video generators (Tora, VACE, WanMove). No public code specified.
• iAm.md: Markdown standard for robot skill self-assessment using agentic introspection with a Qwen3.6-35B VLM. Project page: https://yurimachine.github.io/iAm.md/.
• MotherTree: Tabular transformer for meta-learning decision tree induction, pretrained on synthetic data. Evaluated on TabArena and OpenML CC-18. No public code specified.
• SPERA: EEG foundation model with Joint-Embedding Predictive Architecture (JEPA) and Legendre polynomial spherical priors. Pretrained on 80,000 hours of EEG from 106 datasets. Code: https://github.com/sees4321/SPERA_EEG.
• DIALER: System for rare-class misclassification in edge video analytics using lightweight feature adapters and refinement models on DINOv3-S, SigLIP2-B, MAWS-B. No public code specified.
• EmbodiedRSI: Self-evolving agentic harness for robot learning with a Fast-Slow Dual-System Architecture. Achieves SOTA on RoboCasa365 and LIBERO-Pro with SmolVLA. Project website: https://geeksongs.github.io/agentic_robotics/.
• RoboQuest: Benchmark for goal-directed embodied exploration with 10 tasks (Search, Inspect, Test). Evaluates GPT-6 Astra, Claude Opus 5.5, etc. Code: https://github.com/robocurve/inspect-robots. Project page: https://declare-lab.github.io/RoboQuest/.
• SAPI (Sustainability-Aware Performance Index): Evaluates cell/nucleus segmentation models on performance, energy, and size across 6 CellBinDB datasets. No public code specified.
• RETRO: Tabular foundation model with retrospective inference using Attention Residuals and Gated Attention. Achieves top-three rankings on TabArena, TALENT, and RelArena. No public code specified.
• One-Shot Adaptive Segmentation: Training-free framework for scientific image segmentation using DINOv3 and SAM. No public code specified.
• MorphCL: Self-supervised pretraining for inertial-based Human Activity Recognition (HAR) using motif discovery and feature clustering. No public code specified.
• RoboSteer: Benchmark for behavioral steerability in Behavior Foundation Models (BFMs), evaluating 9 existing BFMs. Project page: https://robosteer.github.io/.
• Argos: Feedforward model for joint scene change detection and 3D reconstruction leveraging Geometric Foundation Models (GFMs) like VGGT-SLAM. Introduces Argos-CD benchmark. No public code specified.
• HySPE (Hyperbolic Symplectic Positional Encoding): Novel positional encoding for length extrapolation using symmetric dual shears. Evaluated on TinyShakespeare and WikiText-103. No public code specified.
• Timer-M1: Multivariate time series foundation model learning from temporal and relational primitives, pretrained on forecasting episodes. Ranks high on FEV, TIME, GIFT-Eval. No public code specified.
• Jev: A decision model that generates calibrated typed answers, used as a reference policy, exploration judge, and replay rater in RL. Evaluated on MiniGrid and Atari games. No public code specified.
• PointVGGT: Zero-shot multiview RGB-D point cloud registration leveraging Visual Geometry Foundation Models (VGGT, Pi3, DepthAnything3). Code: https://github.com/your-repo/pointvggt.
• MINE ODYSSEY: Benchmark for agentic spatial intelligence with 180 navigation tasks in Minecraft reconstructions of real-world locations. Evaluates GPT-6 Astra, Claude Opus 5.5, etc. Code: https://github.com/hkust-nlp/MineOdyssey. Project page: https://mine-odyssey.github.io/.
• RideBench: Large-scale exogenous-aware benchmark for ride-hailing time series forecasting with 4 years of DiDi data across 200 areas. Code: https://github.com/ACAT-SCUT/ridebench.
• Test-Time Compute for Tabular Foundation Models: Investigates adaptation, aggregation, and context construction for TFMs. Introduces DiagScale. Code: https://github.com/kanghui-learning/test-time-compute-for-tabular-foundation-models.
• MiMo-V2.6: Omni-modal LLM family for scaling RL with self-improvement, using hybrid sparse MoE. Open-source release of MiMo-V2.6-Distill-Qwen-9B model, RL framework, and environments: https://mimo.xiaomi.com/rl/mimo-v26.
• MIDAPN: Unified spatiotemporal forecasting backbone for multivariate and multimodal time series, using Multimedia Identity-Aware Graph (MIDAG) and Spectral Prism Convolution (SPConv). Code: https://github.com/MIDAPN.
• FreeInpaint: Feed-forward framework for 3D scene inpainting from unposed, masked image collections, adapting DA3. Project page: https://rorisis.github.io/FreeInpaint/.
• RouterInterp: Interprets MoE expert routing using Sparse Autoencoders to identify features predictive of routing. Code: https://github.com/ilyalasy/routerinterp. Hugging Face: https://huggingface.co/andyrdt/saes-gpt-oss-20b.
• HySPE (Hyperbolic Symplectic Positional Encoding): Novel positional encoding for length extrapolation with 16x zero-shot extrapolation. Code not specified.
• AdaCast: Conditional parameter generation for adaptive time series forecasting using lightweight LoRA updates on a frozen Chronos-Bolt backbone. No public code specified.
• OrionMoE-v1.0: Compact tabular foundation model (37.7M params) with dataset-conditioned sparse Mixture-of-Experts. Code: https://github.com/Lexsi-Labs/OrionMoE. Hugging Face: https://huggingface.co/Lexsi/OrionMoE.
• HART (Hybrid Attention-RWKV Time series Foundation Model): Combines RWKV recurrent blocks with Transformer attention for non-autoregressive multi-step forecasting. Achieves better long-horizon performance than pure RWKV. No public code specified.
• SAM 2.1: Fine-tuned on the HBD building dataset for multi-class building component segmentation, integrated with GARField for 3D reconstruction. No public code specified.
• Verifier-Mediated Multi-LLM Coordination: For multi-robot path planning in CARLA and on Crazyflie aerial robots. No public code specified.
• Zebra3D: Efficient Hybrid Foundation Models for Scalable 3D Reconstruction, converting pretrained models (VGGT, HunyuanWorld-Mirror, π3) into sparse-linear hybrids. Code to be released.
• Mesh-Native Tokenisation (MNC): Retrofits grid-pretrained physics foundation models to unstructured mesh data. Uses Walrus and Overtone. GitHub repository and Hugging Face models to be released.
• OMNI family: Pathology vision encoders (7M-330M parameters) distilled from 10 state-of-the-art foundation models via multi-teacher distillation.
• GlucoPRISM: Protocol-supervised pretraining for cross-cohort glucose transfer using CGM day representations. Uses GlucoFM backbone and various public CGM cohorts. Code: https://anonymous.4open.science/r/glucoprism-4762.
• GOTT: Object-centric dexterous manipulation with a reusable cross-embodiment primitive for multi-fingered robots. Project page: https://dex-gott.github.io.

Impact & The Road Ahead

These advancements demonstrate a clear trajectory towards more capable, efficient, and versatile foundation models. The impact is far-reaching, from democratizing complex AI capabilities by making them accessible on edge devices and with minimal data (e.g., DeFM: Learning Foundation Representations from Depth for Robotics, SurvICL: a Foundation Model for in-context Survival Analysis), to revolutionizing scientific and industrial applications in power grids (Learning by Examples: Power Grid Steady State Analysis with a Tabular Foundation Model), wireless communication (LWM-TF: A Wireless Foundation Model on 5G NR Time-Frequency Resource Grids), and digital pathology (GDBA-Net: A Gated Dual-Branch Adaptation of SAM2 for Robust Cross-Domain Nuclei Segmentation).

The move towards agentic AI (ARC: A Reasoning Recipe for Robot Foundation Models, PhySGA: A Physics-Space Grounded Agentic Framework for Generalizable and Interactive 3D Composition, Mental-Models for Multi-Agent Systems) signals a shift from models that merely predict to systems that can reason, adapt, and interact intelligently, bringing us closer to truly autonomous robots and intelligent assistants. However, challenges remain in ensuring robust generalization across domains (From Adults to Children: Benchmarking Foundation Models for Chest X-Ray Classification under Domain Shift) and managing the computational cost of these powerful models (Tabular Foundation Models Are Effectively Shallow, Physics-Aware Distillation of a Grid Foundation Model for Real-Time Edge Inference).

The road ahead will likely involve further innovation in:

• Adaptive and Conditional Compute: Dynamically adjusting computational resources based on input complexity (e.g., Test-Time Compute for Tabular Foundation Models: Mechanisms, Gains, and Limits, How Much Computation Does Every Forecast Need? Structured Block Skipping in Two Time-Series Foundation Models).
• Reliability and Interpretability: Moving beyond accuracy to evaluate and understand model behavior, especially for critical applications like healthcare (TabReliBench: Evaluating Tabular Foundation Models Beyond Predictive Accuracy, Strategic Governance of AI Models in Earth Science).
• Multimodal Integration: Developing frameworks that seamlessly combine information from diverse modalities for richer world understanding and generation (MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal Data, Kandinsky S2V 1.0: Audio-Driven Avatar Video Generation).
• Ethical AI and Safety: Addressing concerns like data poisoning in protein design systems (Poisoning the corpora of protein design systems during pre-training and retrieval) and ensuring models adhere to complex engineering constraints (ANCHOR: Assessing Naturalistic Constraint Hierarchies and Ordering in Repository-Grounded Agentic Coding).

The rapid evolution of foundation models promises an exciting future where AI systems are not just intelligent but also adaptable, reliable, and deeply integrated into our physical and digital worlds. The ongoing research paves the way for a new generation of AI that learns more efficiently, reasons more profoundly, and interacts more naturally.

Please leave this field empty

Hi there 👋

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

Check your inbox or spam folder to confirm your subscription.