Foundation Models Unleashed: From Drug Discovery to Urban Intelligence, AI’s New Horizons

Latest 50 papers on foundation models: Nov. 16, 2025

The world of AI/ML is buzzing with the transformative power of foundation models, monumental architectures pre-trained on vast datasets that offer unparalleled generalization capabilities. These models are not just about scale; they represent a paradigm shift, enabling rapid adaptation to new tasks and domains with remarkable efficiency. Yet, harnessing their full potential often involves navigating complex challenges, from data scarcity and privacy concerns to maintaining robustness and interpretability. Recent breakthroughs, as highlighted by a collection of cutting-edge research, are pushing the boundaries of what’s possible, extending the reach of foundation models into diverse, high-impact applications.

The Big Idea(s) & Core Innovations:

This wave of innovation spans multiple domains, unified by the strategic application and enhancement of foundation models. In drug discovery, Terray Therapeutics researchers, in their paper “Pretrained Joint Predictions for Scalable Batch Bayesian Optimization of Molecular Designs”, are accelerating the process by integrating pretrained prior functions into Epistemic Neural Networks (ENNs). This significantly boosts the efficiency and accuracy of Batch Bayesian Optimization for molecular design, enabling rapid sampling from joint predictive distributions—a critical step in optimizing large-scale drug discovery. Similarly, the biomedical field is seeing a surge in advanced models. “vMFCoOp: Towards Equilibrium on a Unified Hyperspherical Manifold for Prompting Biomedical VLMs” by authors from Durham University and Tsinghua University, introduces vMFCoOp, a framework that aligns semantic biases in Vision-Language Models (VLMs) and LLMs on a hyperspherical manifold, enhancing few-shot learning and clinical applicability across diverse medical modalities. Another significant leap in medical imaging comes from the Technical University of Munich and King’s College London with “TomoGraphView: 3D Medical Image Classification with Omnidirectional Slice Representations and Graph Neural Networks”. TomoGraphView introduces omnidirectional volume slicing and spherical graph-based feature aggregation, outperforming traditional methods in 3D medical image classification by better capturing spatial relationships.

Computer vision is experiencing multifaceted advancements. The “OmniVGGT: Omni-Modality Driven Visual Geometry Grounded Transformer” by researchers from HKUST, NTU, and Alibaba Group, leverages multiple geometric modalities (depth, camera intrinsics/extrinsics) for superior 3D reconstruction and robotic manipulation. Their GeoAdapter injects geometric information without disrupting the foundation model’s representation space, ensuring stable training. In e-commerce, the Ohio State University team, in their work “Captions Speak Louder than Images: Generalizing Foundation Models for E-commerce from High-quality Multimodal Instruction Data”, introduced MMECInstruct and CASLIE, a lightweight framework for multimodal understanding, demonstrating the power of high-quality multimodal data for better product insights. Addressing crucial ethical concerns, “Privacy Beyond Pixels: Latent Anonymization for Privacy-Preserving Video Understanding” by the University of Central Florida proposes SPLAVU, a novel method for anonymizing latent features in video, significantly reducing privacy leakage while maintaining task performance. For time series analysis, papers like “Spectral Predictability as a Fast Reliability Indicator for Time Series Forecasting Model Selection” from UCLA introduce spectral predictability (ℙ) as a crucial metric for efficient model selection, revealing that large Time Series Foundation Models (TSFMs) excel in high-predictability datasets. Building on this, “Are Time-Indexed Foundation Models the Future of Time Series Imputation?” by EDF R&D demonstrates the robust zero-shot imputation capabilities of models like TabPFN-TS and MoTM across diverse datasets.

Under the Hood: Models, Datasets, & Benchmarks:

These innovations are powered by new models, enhanced architectures, and meticulously crafted datasets and benchmarks:

• TransactionGPT (“TransactionGPT” by Visa Research): A novel 3D-Transformer architecture with a virtual token mechanism for understanding and generating complex consumer transaction data, achieving a 22% improvement in classification over production models.
• OmniVGGT (“OmniVGGT: Omni-Modality Driven Visual Geometry Grounded Transformer” by HKUST, NTU, SYSU, NUS, Alibaba Group): Features the GeoAdapter for stable geometric information injection and a stochastic multimodal fusion strategy for robust spatial representations.
• MMECInstruct Dataset & CASLIE Framework (“Captions Speak Louder than Images: Generalizing Foundation Models for E-commerce from High-quality Multimodal Instruction Data” by The Ohio State University): MMECInstruct is the first high-quality multimodal instruction dataset for e-commerce, enabling the lightweight CASLIE framework to generalize across diverse applications. Code available at https://ninglab.github.io/CASLIE/.
• DSANet (“Learning to Tell Apart: Weakly Supervised Video Anomaly Detection via Disentangled Semantic Alignment” by Huazhong University of Science and Technology): A Disentangled Semantic Alignment Network for weakly supervised video anomaly detection, using self-guided normality modeling and decoupled contrastive semantic alignment. Code available at https://github.com/lessiYin/DSANet.
• EEG-X (“EEG-X: Device-Agnostic and Noise-Robust Foundation Model for EEG” by Emotiv Research and Monash University): A device-agnostic and noise-robust foundation model for EEG analysis, featuring location-based channel embeddings and noise-aware reconstruction strategies. Code available at https://github.com/Emotiv/EEG-X.
• LandSegmenter (“LandSegmenter: Towards a Flexible Foundation Model for Land Use and Land Cover Mapping” by Technical University of Munich): The first LULC foundation model, using weak supervision and confidence-guided fusion strategies for flexible and accurate land use and land cover mapping. Code available at https://github.com/zhu-xlab/LandSegmenter.git.
• BuildingWorld Dataset & Cyber City Generator (“BuildingWorld: A Structured 3D Building Dataset for Urban Foundation Models” by University of Calgary and Shenzhen University): The largest structured 3D building dataset with five million LOD2 models, coupled with a virtual city generator to enhance urban foundation models. Utilizes Helios++ LiDAR Simulator (code: https://github.com/helios-ml/helios).
• GRAVER (“GRAVER: Generative Graph Vocabularies for Robust Graph Foundation Models Fine-tuning” by Beihang University, Guangxi Normal University, National University of Singapore, and University of Illinois, Chicago): Enhances robustness and efficiency of fine-tuning Graph Foundation Models (GFMs) through generative graph vocabularies and a MoE-CoE routing mechanism. Code available at https://github.com/RingBDStack/GRAVER.
• xBD-S12 Dataset (“The Potential of Copernicus Satellites for Disaster Response: Retrieving Building Damage from Sentinel-1 and Sentinel-2” by ETH Zurich and University of Zurich): A novel dataset aligning xBD image pairs with Sentinel-1 and Sentinel-2 acquisitions for building damage assessment. Code available at https://github.com/olidietrich/xbd-s12.
• HISTOPANTUM Dataset & HistoDomainBed Framework (“Benchmarking Domain Generalization Algorithms in Computational Pathology” by University of Warwick, Histofy Ltd, and ICMR-National Institute of Pathology): A large-scale tumor patch dataset for pan-cancer tumor detection, along with a benchmarking framework for domain generalization algorithms in computational pathology. Code available at https://github.com/mostafajahanifar/HistoDomainBed.

Impact & The Road Ahead:

These advancements signify a pivotal moment for foundation models, pushing them beyond general-purpose tasks into specialized, high-stakes applications. The potential impact is enormous: faster drug discovery with optimized molecular designs, more adaptable robots capable of understanding natural language instructions through “VLAD-Grasp: Zero-shot Grasp Detection via Vision-Language Models” by University of Washington, Google Research, Stanford University, ETH Zurich, and MIT CSAIL, and more robust medical diagnostics with precise 3D image classification. The move towards decentralized, blockchain-secured RAG systems, as seen in “A Decentralized Retrieval Augmented Generation System with Source Reliabilities Secured on Blockchain” by the University of Notre Dame, promises enhanced transparency and trustworthiness in AI-driven information retrieval. Furthermore, the focus on bias mitigation through methods like ForAug (“ForAug: Recombining Foregrounds and Backgrounds to Improve Vision Transformer Training with Bias Mitigation” by RPTU University Kaiserslautern-Landau and German Research Center for Artificial Intelligence) and ethical considerations in video understanding with SPLAVU underscores a growing commitment to responsible AI development.

Looking ahead, we’re seeing the emergence of true world models, capable of simulating complex visual dynamics and interactions, as surveyed in “Simulating the Visual World with Artificial Intelligence: A Roadmap” by Carnegie Mellon University, Nanyang Technological University, and Kuaishou Technology. This, coupled with the pursuit of Artificial General Intelligence (AGI) through frameworks like the “Intelligence Foundation Model: A New Perspective to Approach Artificial General Intelligence” from Tsinghua University, points to a future where AI systems possess deeper cognitive abilities and a more nuanced understanding of the world. The challenges, such as spectral shift in time series models highlighted by “Frequency Matters: When Time Series Foundation Models Fail Under Spectral Shift” from King AI Labs/Microsoft Gaming and KTH Royal Institute of Technology, remind us that fine-tuning and domain adaptation remain crucial. Yet, with continued research into flexible architectures, robust data strategies, and ethical considerations, foundation models are poised to revolutionize nearly every facet of our technological landscape.