Foundation Models Unleashed: From Humanoid Control and 8K Sensing to Trustworthy Medical AI

Latest 50 papers on foundation models: Nov. 10, 2025

The landscape of AI/ML is being rapidly reshaped by Foundation Models (FMs), which are not only growing in scale but are also being expertly adapted to solve high-stakes, domain-specific challenges across robotics, medicine, and scientific discovery. These models are moving beyond general-purpose tasks to become hyper-specialized tools that maintain generalization power while achieving state-of-the-art performance in complex, constrained environments.

The Big Idea(s) & Core Innovations

Recent research underscores a dual theme: developing highly specialized FMs for real-world impact and enhancing the efficiency and trustworthiness of their deployment.

1. Specialization and Real-World Autonomy: A major focus is equipping FMs with dynamic, real-world reasoning. In robotics, researchers from Carnegie Mellon University and Meta introduced BFM-Zero in their work, BFM-Zero: A Promptable Behavioral Foundation Model for Humanoid Control Using Unsupervised Reinforcement Learning. This breakthrough enables humanoid robots to execute diverse tasks via promptable generalist policies without retraining, leveraging unsupervised reinforcement learning to bridge the sim-to-real gap. Similarly, the paper UniLION (UniLION: Towards Unified Autonomous Driving Model with Linear Group RNNs) offers a unified framework for autonomous driving, eliminating the need for explicit fusion modules by integrating multi-modal and temporal data through a shared 3D backbone.

2. Trust and Efficiency in Specialized Domains: The adaptation of FMs for fields like medical imaging and remote sensing requires addressing challenges of data scarcity, domain shift, and reliability. This is seen in PLUTO-4 (PLUTO-4: Frontier Pathology Foundation Models) from PathAI, which introduces Vision Transformer models pretrained on over 500k whole slide images, achieving robust generalization across disease types. Complementing this, papers like FusionDP (FusionDP: Foundation Model-Assisted Differentially Private Learning for Partially Sensitive Features) enhance privacy by selectively protecting only sensitive features using FMs for imputation, improving the privacy-utility trade-off critical for clinical data.

3. Scaling and Unification for Unstructured Data: Advances are also emerging in handling complex, unstructured data. In high-energy physics, the creation of Aspen Open Jets (Aspen Open Jets: Unlocking LHC Data for Foundation Models in Particle Physics), the largest ML-ready dataset for LHC, enables the pre-training of jet-based FMs. For time series, models like Datadog’s TOTO (from This Time is Different: An Observability Perspective on Time Series Foundation Models) and Tsinghua University’s Sundial (Sundial: A Family of Highly Capable Time Series Foundation Models) achieve state-of-the-art zero-shot forecasting by optimizing for observability data and using novel continuous tokenization techniques, respectively.

Under the Hood: Models, Datasets, & Benchmarks

The ability to create specialized FMs relies heavily on large, high-quality, and often domain-specific resources:

• GeoLLaVA-8K (GeoLLaVA-8K: Scaling Remote-Sensing Multimodal Large Language Models to 8K Resolution) introduced SuperRS-VQA (average 8K resolution) and HighRS-VQA to enable UHR remote sensing analysis.
• NABench (NABench: Large-Scale Benchmarks of Nucleotide Foundation Models for Fitness Prediction) aggregates over 2.6 million mutated sequences, providing the largest and most comprehensive benchmark for Nucleotide Foundation Models (NFMs).
• IMO-Bench (Towards Robust Mathematical Reasoning) offers a suite of benchmarks focusing on rigorous, multi-step mathematical reasoning, complete with automated graders.
• FetalUS-188K (introduced in Challenging DINOv3 Foundation Model under Low Inter-Class Variability: A Case Study on Fetal Brain Ultrasound) provides a new multi-center dataset to test FM robustness in medical scenarios characterized by low inter-class variability.
• TOTO and BOOM: TOTO is a 151-million-parameter time series FM, optimized and benchmarked using BOOM, the first large-scale benchmark for observability metrics.
• PLLuM (PLLuM: A Family of Polish Large Language Models) constructed a massive 140-billion-token Polish text corpus to support culturally relevant AI beyond English-centric systems.

Developers looking to leverage these advancements should explore public repositories, such as the code for the lightweight tabular FM nanoTabPFN (nanoTabPFN: A Lightweight and Educational Reimplementation of TabPFN), which is available on GitHub, or the unified tabular framework TabTune (TabTune: A Unified Library for Inference and Fine-Tuning Tabular Foundation Models).

Impact & The Road Ahead

These advancements point toward a future where FMs are modular, efficient, and domain-aware. Efficiency is paramount: research in Revisiting Federated Fine-Tuning: A Single Communication Round is Enough for Foundation Models demonstrates that federated fine-tuning can be effective with a single communication round, drastically cutting network overhead for distributed training.

The rise of multi-agent and orchestration frameworks, such as Agent-Omni (Agent-Omni: Test-Time Multimodal Reasoning via Model Coordination for Understanding Anything) and the multi-agent system for medical pre-consultation (From Passive to Proactive: A Multi-Agent System with Dynamic Task Orchestration for Intelligent Medical Pre-Consultation), suggests a shift from monolithic models to coordinated systems of specialized FMs.

However, challenges remain. The paper When One Modality Sabotages the Others: A Diagnostic Lens on Multimodal Reasoning warns of ‘modality sabotage,’ where one data stream can undermine the entire prediction—underscoring the ongoing need for robust diagnostic tools. Similarly, How Far Are Surgeons from Surgical World Models? shows that while generative models can produce photorealistic surgical videos, they lack the deep causal logic necessary for true ‘world model’ simulation.

Ultimately, the path forward involves rigorous benchmarking (as provided by NABench and IMO-Bench), improved efficiency (via single-round federated tuning and prompt-expert mixtures like GMoPE), and domain-specific adaptation to ensure that the power of foundation models translates into reliable, trustworthy, and actionable intelligence across every specialized domain.