From Pixels to PDEs: Unveiling the Next Generation of Foundation Models

Latest 50 papers on foundation models: Sep. 29, 2025

Foundation models are rapidly transforming the AI landscape, demonstrating unprecedented capabilities across diverse domains. From understanding complex medical images to predicting dynamic land-surface processes and even reasoning about cosmic phenomena, these powerful models are proving to be much more than just large language processors. This digest dives into a collection of recent research papers, showcasing breakthroughs, innovative architectures, and novel applications that are pushing the boundaries of what foundation models can achieve.### The Big Idea(s) & Core Innovationsthe heart of these advancements lies a common thread: tackling complex, data-rich problems with scalable and often self-supervised approaches. Many papers focus on enhancing existing foundation models or developing new ones tailored to specific, challenging domains. For instance, in medical imaging, GE Healthcare’s “Decipher-MR: A Vision-Language Foundation Model for 3D MRI Representations” introduces a 3D MRI-specific vision-language foundation model trained on a massive dataset of over 200,000 MRI series. This model excels in diverse clinical tasks like disease classification and cross-modal retrieval, proving robust across anatomical regions., for mammogram interpretation, a team from the Hong Kong University of Science and Technology and Sun Yat-sen Memorial Hospital presented “A Versatile Foundation Model for AI-enabled Mammogram Interpretation”, VersaMammo. This model leverages a two-stage pre-training strategy and the largest mammogram dataset to date, achieving state-of-the-art results across 92 specific clinical tasks.medical applications, the concept of Visual Instruction Pretraining (ViTP), introduced by researchers from Nankai University, in “Visual Instruction Pretraining for Domain-Specific Foundation Models”, marks a significant shift. ViTP integrates high-level reasoning directly into vision backbone learning for domain-specific tasks like remote sensing, enhancing low-level perception. This ‘top-down’ approach, paired with Visual Robustness Learning (VRL), allows for more semantically rich and robust visual features.innovative approach comes from the Pohang University of Science and Technology, with “MOMEMTO: Patch-based Memory Gate Model in Time Series Foundation Model”. MOMEMTO tackles over-generalization in time series anomaly detection by using a patch-based memory gate module that stores representative normal patterns from multiple domains. This is crucial for efficient and accurate few-shot learning.the realm of scientific machine learning, Zituo Chen and Sili Deng from MIT introduce “Flow marching for a generative PDE foundation model”. This groundbreaking algorithm unifies neural operator learning with flow matching to create a generative PDE foundation model capable of uncertainty-aware ensemble generation and stable long-term predictions. This moves beyond traditional deterministic modeling, offering a more comprehensive understanding of complex dynamical systems., the World Bank Group demonstrated the practical application of AI in “AI-Derived Structural Building Intelligence for Urban Resilience: An Application in Saint Vincent and the Grenadines”, showing how fine-tuned deep learning models outperform general geospatial foundation models for rooftop classification in small island developing states.foundational challenges, Dujin Lee et al. from Korea University in “Training-Free Label Space Alignment for Universal Domain Adaptation” propose Training-free Label Space Alignment (TLSA), leveraging VLMs like CLIP to align label spaces for Universal Domain Adaptation (UniDA), achieving significant performance gains without extensive training. This highlights a shift from visual feature alignment to more abstract label space alignment.### Under the Hood: Models, Datasets, & Benchmarksinnovations discussed are often underpinned by novel architectural designs, large-scale datasets, and new benchmarking methodologies:Decipher-MR (https://arxiv.org/pdf/2509.21249) is a 3D MRI-specific vision-language foundation model trained on 200,000+ MRI series from 22,000 studies. Code available at https://github.com/gehealthcare/Decipher-MR and https://huggingface.co/gehealthcare/decipher-mr.VersaMammo (https://arxiv.org/pdf/2509.20271) utilized a curated dataset of 706,239 mammogram images from 21 multi-institutional sources and introduced a comprehensive benchmark for 92 specific clinical tasks.StefaLand (https://arxiv.org/pdf/2509.17942), from Pennsylvania State University, is a geoscience foundation model using attribute-centric pretraining for dynamic land-surface predictions. Code and datasets are publicly released at https://anonymous.4open.science/r/StefaLand-9421/.MOMEMTO (https://arxiv.org/pdf/2509.18751) uses a patch-based memory gate module and multi-domain training, specializing in time series anomaly detection, showing strong results on 23 univariate benchmark datasets.The generative PDE foundation model (https://arxiv.org/abs/2509.18611) by Chen Shen and Stefan Ringeisen et al. from MIT introduces an extensive heterogeneous PDE corpus (2.5 million trajectories, 233GB) for pretraining. Code available at https://github.com/zituo-chen/flow-marching.SoM-1K (https://som-1k.github.io/), introduced by Qixin Wan et al., is the first large-scale multimodal benchmark for Strength of Materials problems, challenging foundation models in visual-textual reasoning. The dataset is available at https://som-1k.github.io/.CaTS-Bench (https://huggingface.co/datasets/a9f3c7e2/CaTSBench), by Luca Zhou et al. from Sapienza University of Rome, is the first large-scale benchmark for context-aware time series captioning and reasoning, including rich metadata and visual plots. It is available on Hugging Face.GraphUniverse (https://graphuniverse.streamlit.app/), from Universitat Politècnica de Catalunya and UC Santa Barbara, offers a framework for systematic evaluation of inductive generalization in graph learning with a hierarchical generative model for graph families. Code available via PyPi: https://pypi.org/project/graphuniverse/.OpenGVL (https://arxiv.org/abs/2509.17321), a benchmark for Visual Temporal Progress in robotics, developed by Y. J. Ma et al. from various institutions including Google Research and Hugging Face, helps curate large-scale open-source datasets by evaluating Vision-Language-Action (VLA) models on temporal task progress prediction. Code is available at https://github.com/AlexanderKoch-Koch/low.MolPILE (https://huggingface.co/datasets/scikit-fingerprints/MolPILE), introduced by Jakub Adamczyk et al. from AGH University of Krakow, is a large-scale (222 million compounds) and diverse dataset for molecular representation learning. Code is available at https://github.com/scikit-fingerprints/MolPILE_dataset.### Impact & The Road Aheadpapers collectively paint a picture of foundation models evolving beyond mere prediction towards deeper understanding and more robust generalization across modalities and domains. The shift towards self-supervised learning, cross-modal knowledge transfer, and domain-specific adaptation, as seen in IBM Research Europe’s “A Sentinel-3 Foundation Model for Ocean Colour”, promises to unlock powerful new capabilities in fields from environmental monitoring to industrial automation.are seeing a clear trend where large language models are even surpassing domain-specific architectures, as highlighted by Sheng Wong et al. from Oxford Digital Health Labs in their paper “Large language models surpass domain-specific architectures for antepartum electronic fetal monitoring analysis”. This signals a future where generalist models, fine-tuned or adapted, can achieve expert-level performance in highly specialized tasks.remain, such as the “knowing-doing gap” in reinforcement learning, where foundation models struggle with low-level control despite high-level reasoning, as discussed by Remo Sasso et al. from Queen Mary University of London in “Exploration with Foundation Models: Capabilities, Limitations, and Hybrid Approaches”. Also, the crucial need for interpretability in high-stakes domains like brain science and medical diagnosis is underscored by Thomas Serre and Ellie Pavlick from Brown University in “From Prediction to Understanding: Will AI Foundation Models Transform Brain Science?”., the rapid progress, demonstrated by models like Google Research’s Veo 3 in “Video models are zero-shot learners and reasoners” showing zero-shot reasoning in visual tasks, points towards a future where foundation models become truly versatile, general-purpose AI agents. The ongoing development of efficient adaptation methods like HyperAdapt (https://arxiv.org/pdf/2509.18629) from Purdue University, and DEFLECT (https://arxiv.org/pdf/2503.09493) from CNES and European Space Agency Φ-Lab, will enable these models to be deployed more widely and cost-effectively, bringing us closer to a future where AI empowers discovery and innovation across every scientific and industrial frontier.