Foundation Models: Charting New Territories from Neurons to Oncology

Latest 50 papers on foundation models: Sep. 1, 2025

The world of AI is rapidly expanding, with Foundation Models (FMs) at the forefront of this revolution. These massive, pre-trained models are reshaping how we approach complex problems across diverse fields, from unraveling the mysteries of the human brain to accelerating medical diagnostics. However, adapting these generalist powerhouses to specialized tasks, ensuring their robustness, and understanding their ethical implications remain critical challenges. Recent research has been pushing these boundaries, revealing exciting breakthroughs and practical applications.

The Big Idea(s) & Core Innovations

At the heart of many recent advancements lies the ingenious adaptation of large-scale pre-trained models to domain-specific challenges, often by marrying them with novel architectural components or fine-tuning strategies. For instance, in medical imaging, we see a powerful trend of leveraging high-fidelity features from FMs. Researchers from the University of Paris-Saclay, INRIA, and Google Research introduce Dino U-Net: Exploiting High-Fidelity Dense Features from Foundation Models for Medical Image Segmentation, demonstrating that integrating self-supervised pre-training with traditional U-Net architectures significantly boosts accuracy and robustness in medical image segmentation. Similarly, ShanghaiTech University’s Adapting Foundation Model for Dental Caries Detection with Dual-View Co-Training proposes DVCTNet, which mimics clinical workflows by fusing global and local dental image views using a Gated Cross-View Attention mechanism to achieve superior caries detection. These efforts underscore the potential for FMs to bring precision and reliability to critical healthcare applications.

Beyond vision, the versatility of FMs is being explored in unexpected domains. HSE University and Yandex Research present Turning Tabular Foundation Models into Graph Foundation Models, showcasing G2T-FM, a framework that ingeniously transforms graph tasks into tabular ones using TabPFNv2. This opens up an entirely new avenue for graph machine learning, proving the generalizability of FMs across data structures. The fusion concept extends to multimodal understanding; in Towards Open-Vocabulary Multimodal 3D Object Detection with Attributes, researchers from the University of California, Davis and Mitsubishi Electric Research Laboratories (MERL) introduce OVODA, enabling open-vocabulary 3D object and attribute detection without needing prior knowledge of novel class anchor sizes – a significant leap for autonomous systems. Meanwhile, Sirui Li and Northwestern University’s TAGS: 3D Tumor-Adaptive Guidance for SAM creatively adapts the Segment Anything Model (SAM) for 3D tumor segmentation, leveraging CLIP’s semantic insights and multi-prompt fusion to achieve state-of-the-art results with minimal parameters.

The ethical and practical implications of FMs are also under intense scrutiny. The Centre for Technomoral Futures and Gifford Scholarship in AI Ethics delve into Technology as uncharted territory: Contextual integrity and the notion of AI as new ethical ground, urging a re-evaluation of AI ethics within sociotechnical contexts. This call for deeper understanding is echoed by Fraunhofer Heinrich Hertz Institute and Technical University of Berlin in Model Science: getting serious about verification, explanation and control of AI systems, proposing a new field focused on ensuring the safety, transparency, and reliability of complex AI systems.

Under the Hood: Models, Datasets, & Benchmarks

The innovations discussed rely heavily on specialized models, datasets, and rigorous benchmarking, which are crucial for evaluating and advancing the field:

• Dino U-Net & DVCTNet: These medical segmentation frameworks leverage the foundational strengths of DINOv2 and ViTAdapter, demonstrating the power of pre-trained vision models. DVCTNet also introduces the first high-precision benchmark dataset for dental caries detection, double-verified through intra-oral images and panoramic X-rays.
• G2T-FM: This approach utilizes TabPFNv2 as its backbone, showcasing the adaptability of tabular foundation models to new domains.
• TAGS: Adapts the Segment Anything Model (SAM) and integrates CLIP’s semantic embeddings for 3D medical tasks.
• OVODA: Combines foundation model features with prompt tuning and introduces the OVAD dataset for attribute detection in 3D scenes.
• HoneyBee: A modular framework for oncology that integrates diverse biomedical modalities (clinical text, radiology, pathology, molecular profiles) from The Cancer Genome Atlas (TCGA) and NCI Cancer Research Data Commons (CRDC). Code available at HoneyBee and Hugging Face.
• MMTU: A Massive Multi-Task Table Understanding and Reasoning Benchmark with over 30K questions across 25 real-world table tasks. Code at MMTU GitHub and Hugging Face.
• EEG-FM-Bench: The first comprehensive benchmark for systematic evaluation of EEG foundation models, providing extensive baselines for comparison. Code available at EEG-FM-Bench GitHub.
• SegEarth-OV: An annotation-free open-vocabulary segmentation framework for remote sensing, featuring SimFeatUp and Global Bias Alleviation. Code at SegEarth-OV-2 GitHub.
• LodeStar: Utilizes foundation models for automatic skill segmentation in robotics, with code for xArm-Developer at GitHub.
• ArtFace: Fine-tunes the vision–language foundation model CLIP with LoRA and complements it with embeddings from an adapted face recognition network. Resources at ArtFace.
• EQUATE: A fine-tuning framework for symbolic regression via foundation model distillation. Code available at EQUATE GitHub and Hugging Face Space.

Impact & The Road Ahead

The impact of these advancements is profound, promising more accurate medical diagnoses, smarter autonomous systems, and more robust AI in general. For medical AI, the development of models like OmniMRI by Harvard Medical School (OmniMRI: A Unified Vision–Language Foundation Model for Generalist MRI Interpretation) signifies a move towards generalist models that can handle the entire MRI workflow, from reconstruction to report generation. This could fundamentally transform radiology. Similarly, MD Anderson Cancer Center’s The Next Layer: Augmenting Foundation Models with Structure-Preserving and Attention-Guided Learning for Local Patches to Global Context Awareness in Computational Pathology with EAGLE-Net offers more accurate and interpretable tools for computational pathology, aiding clinical decision-making.

In robotics, frameworks like LODESTAR (LodeStar: Long-horizon Dexterity via Synthetic Data Augmentation from Human Demonstrations) from the International Conference on Learning Representations (ICLR) demonstrate how synthetic data and foundation models can enable complex dexterous manipulation with limited real-world data, accelerating the development of adaptable robots. Furthermore, the integration of explainable AI (XAI) and robustness measures, as seen in Borealis AI’s Robustness Feature Adapter for Efficient Adversarial Training, is crucial for building trustworthy AI in high-stakes applications.

The future of foundation models will likely see continued efforts in multi-modal fusion, domain adaptation, and the development of standardized benchmarks to ensure fairness and transparency. The survey by University of Amsterdam on Looking Beyond the Obvious: A Survey on Abstract Concept Recognition for Video Understanding and the discussion on Why Relational Graphs Will Save the Next Generation of Vision Foundation Models? by University of Tehran highlight the ongoing pursuit of more sophisticated reasoning capabilities in vision. As these models become increasingly pervasive, the concurrent evolution of ‘Model Science’ and AI ethics will be paramount to ensure their responsible development and deployment. The journey to truly intelligent, reliable, and ethical AI is just beginning, and foundation models are undeniably leading the way.