Wireless Foundation Models

Deploying foundation models for robot control raises a central challenge: the expressive power that enables rich, multimodal perception also makes these models opaque and difficult to analyze formally, rendering them intractable for existing verification tools. In this paper, we present FEARL (Foundation-Enabled Assured Robot Learning), a framework that addresses this tension through a modular architectural decomposition. FEARL separates the policy into a large Controller (C) responsible for high-dimensional perception and task reasoning, and a small Safety module (S) that receives low-dimensional observations from dedicated safety sensors together with a bounded context embedding from C and produces the final action. Since many robot safety requirements, such as collision avoidance and workspace boundary constraints, can be expressed over these safety sensor observations, formal verification can be applied to S rather than to the full foundation-model backbone. This makes formal analysis tractable with existing tools while preserving the Controller's expressive power for task reasoning. To show that the decomposed policy remains capable of solving diverse tasks, we evaluate FEARL on three simulated robotic domains using multiple Controller backbones and training procedures, including pretrained off-the-shelf vision-language-action models. We further transfer the learned policy from one of our simulated tasks to a physical robot, suggesting that the low-dimensional safety interface supports practical sim-to-real transfer.

Foundation models have achieved remarkable performance across medical question answering, imaging, and electronic health record (EHR) tasks, yet reliable clinical deployment remains challenging due to limited interpretability, vulnerability to distribution shift, and weak alignment with clinician reasoning. We argue that these limitations arise because existing approaches prioritize representation learning while treating patient comparison as an emergent property rather than a primary source of clinical evidence. To address this gap, we propose CAFM, a Cohort-Anchored Foundation Model framework that elevates patient cohorts to a first-class object throughout the learning pipeline. The framework consists of four stages: deviation-aware data curation, cohort-conditioned pretraining, multimodal cohort alignment, and clinician-in-the-loop refinement. Together, these stages improve data quality, organize representations around clinically meaningful cohort structure, preserve modality-specific relationships, and support auditable clinical decision-making. The framework is compositional and can augment existing EHR foundation models without modifying their underlying encoders. We illustrate CAFM through four clinical case studies spanning acute kidney injury prediction, cardiovascular risk stratification from electrocardiograms, optic neuropathy triage from orbital imaging, and electroretinogram-grounded report generation. We further present five empirically testable hypotheses and identify open challenges in data quality, irregular temporality, multimodal learning, distribution shift, and evaluation beyond predictive accuracy. We argue that explicitly anchoring foundation models to patient cohorts provides a principled path toward trustworthy clinical AI.

Vision foundation models have substantially advanced computer vision, enabling state-of-the-art performance in zero- and few-shot settings. They have been successfully applied to biomedical imaging tasks ranging from organ segmentation in computed tomography to cell segmentation in light microscopy. Electron microscopy (EM) is a central modality for analyzing cellular ultrastructure due to its nanometer-scale resolution. However, the application of foundation models in EM has so far been limited to specific organelles, such as mitochondria, largely due to the diversity of segmentation tasks and the scarcity of comprehensively annotated data. As a result, EM segmentation still predominantly relies on supervised learning, requiring extensive manual annotation and limiting ultrastructural analysis. To address this gap, we propose $μ$Match, a framework for semi-supervised learning and domain adaptation that leverages foundation models. We implement state-of-the-art student-teacher-based methods and evaluate multiple foundation models (SAM, SAM2, $μ$SAM, DINOv2/v3) on challenging EM tasks, including mitochondrion, nucleus, and neurite segmentation. Our results demonstrate consistent improvements over strong baselines and highlight a path toward substantially reducing the annotation effort in EM.

We present Casper3D, a lightweight probabilistic framework for converting noisy multi-view 2D foundation-model embeddings into a latent 3D semantic representation. We model view-level semantic features as noisy observations of an underlying 3D semantic state and infer this state with a set-based variational model that incorporates relative pose during multi-view reasoning. Casper3D is trained by predicting held-out semantic observations from novel viewpoints, while remaining aligned with visual and text semantic spaces for open-vocabulary 3D understanding. The framework is backbone-agnostic and applies to both language-aligned and self-supervised embeddings. Experiments show that Casper3D produces more stable 3D semantics than simple multi-view pooling, especially in ambiguous and noisy settings.

Causal discovery is critical for understanding complex data-generating mechanisms, yet traditional algorithms often struggle with highly non-linear and noisy systems, or suffer from severe computational bottlenecks. Recent tabular foundation models based on Prior-Data Fitted Networks (PFNs) have demonstrated remarkable zero-shot inference capabilities, but their potential for explicit structural causal discovery remains underexplored. To bridge this gap, we propose DCD-PFN, a decoupling-aware foundation model for causal discovery. Instead of directly amortizing global graph reconstruction, DCD-PFN focuses on local causal discovery through a decoupling-based paradigm. Through pre-training on diverse synthetic Structural Causal Models (SCMs), the model learns sample-wise decoupling weights that enable Markov boundary (MB) identification. Furthermore, by leveraging parallelized local discovery, DCD-PFN efficiently reconstructs global causal graphs while remaining grounded in the theoretical foundations of decoupling-based causal discovery. Experiments demonstrate that our foundation model achieves robust zero-shot generalization.

We present OmniV2X, a generative foundation model for vehicle-to-everything (V2X) cooperative driving. The model directly interprets independent context sequences comprising multi-modal and multi-agent observations. The new design mitigates the computational cost of dense 3D perception, the vulnerability to data scarcity in cooperative scenarios, and the poor compliance with standardized messaging in existing methods that fuse multi-modal inputs into a shared representation. For training, we present an end-to-end supervised pipeline using a downstream trajectory generation loss, in which a high-capacity generative sequence planner implicitly learns to steer the model and leverage multi-modal inputs via cross-attention injection. As a foundation model, we demonstrate that OmniV2X pre-trained on large-scale single-agent planning datasets can efficiently adapt to cooperative environments by integrating the conditioning context with lightweight, standard-compliant V2X tokens. Evaluated on the DAIR-V2X-Seq dataset, OmniV2X outperforms existing end-to-end cooperative driving baselines, achieving state-of-the-art performance with less than 10% of the fine-tune V2X dataset and less than 1% of the communication bandwidth. We conduct comprehensive evaluations to demonstrate its computational efficiency and robustness under real-world constraints.

EEG foundation models can learn generalizable representations from large-scale EEG corpora to enable single-backbone transfer across diverse clinical and brain-computer interface tasks. Existing models typically discretize the continuous multi-channel EEG waveform into patches or codebook tokens and train a transformer with masked self-supervision. Recognizing that this discretization fragments continuous brain rhythms and obscures fine-grained temporal dynamics, we present B[FM]$^2$(Brain Foundation Model via Flow Matching), whose inductive bias aligns with the data by pretraining directly on the raw signal using continuous-time flow matching without patches, tokenization, or masking. However, multi-channel EEG signals pose an architectural challenge for flow matching: time is densely sampled and highly autocorrelated (thousands of timepoints), while the electrode axis is short (tens of channels) at distinct scalp positions. To address this time-electrode asymmetry, we introduce SplitUNet, a velocity network that factorizes each block into separate 1D temporal and 1D electrode convolutions and downsamples only along time, preserving electrode topology throughout the hierarchy. B[FM]$^2$ sets a new state of the art on 7 of 9 standard downstream EEG classification tasks, using a pretraining budget of only 36,895 segments ($\approx$ 307h), 1-2 orders of magnitude ($\approx$ 30x) less than required by existing EEG foundation models. Further, it generates synthetic EEGs that two board-certified neurologists cannot distinguish from brain data (Cohen's $κ=$ -0.096). https://jd730.github.io/projects/BFM2

Embodied foundation models are expected to benefit from data scaling like large language models, but face a much tighter data bottleneck. Teleoperated real-robot trajectories remain the dominant pretraining source due to their precise action supervision and embodiment alignment, yet their scalability is limited by high collection cost, acquisition difficulty, and low behavioral and environmental diversity. These limitations have sparked interest in egocentric human video as a scalable, substantially lower-cost, and more diverse alternative for embodied model pretraining. However, its effectiveness compared to teleoperated real-robot data remains underexplored. To address this question, we conduct a systematic study comparing egocentric human video and teleoperated real-robot trajectories as pretraining data sources for embodied foundation models, under fixed post-training and validation protocols. Surprisingly, we find that egocentric data, when processed through a carefully designed filtering and labeling pipeline, is not merely a viable substitute for model pretraining but can lead to superior performance. With the same amount of pretraining data, models pretrained on egocentric data achieve a 24% lower validation loss on real-robot action prediction, as well as 52.5% and 90% higher success rates on in-distribution and out-of-distribution real-robot task execution, respectively. This finding verifies a scalable paradigm for embodied foundation models: pretrain on egocentric human video to learn diverse world representations, then adapt with a small amount of labeled real-robot data for action-space alignment. We hope this study encourages broader exploration of egocentric data and offers guidance for data quality assessment before costly robot data collection.

Nuclear Magnetic Resonance (NMR) spectroscopy is a powerful tool for molecular structure analysis, and spectral artificial intelligence offers great potential for its rapid and automated interpretation. However, the scarcity of experimental NMR datasets has constrained deep learning in this domain to narrow, task-specific applications that lack broad generalization. Here, we introduce UltraNMR, a large-scale foundation model for NMR that leverages the intrinsic properties of NMR spectra to learn generalizable spectral representations. We collected 158 million paired simulated $^{1}$H and $^{13}$C NMR spectra to train UltraNMR, employing multiple domain-specific pre-training objectives. UltraNMR captures both intra-spectral and inter-spectral dependencies, enabling seamless simulation-to-real adaptation. We demonstrate that adapting UltraNMR to a range of molecular structure analysis tasks on experimental NMR spectra consistently yields state-of-the-art performance and clearly outperforms UltraNMR variants trained directly on downstream data without simulation pre-training. We also construct a large-scale NMR spectral vector library by encoding simulated NMR spectra using UltraNMR, covering 94 million unique molecules and enabling effective structure-aware retrieval. In real-world applications, UltraNMR facilitates the structural elucidation of two previously unknown natural products from Chinese herbal medicines recorded in the Chinese Pharmacopoeia. These results suggest that large-scale simulation pre-training can effectively bridge the simulation-to-real gap, enabling robust and generalizable molecular structure analysis of real-world NMR spectra.

We introduce the first generative foundation model for chest radiograph synthesis trained from scratch at the billion-parameter scale. Existing radiographic AI models often suffer from poor generalisation across patient subpopulations, institutions, and acquisition settings, resulting in limited real-world clinical utility. Controlled, high-fidelity synthesis of chest radiographs is a promising path toward diversifying clinical datasets and evaluating the robustness of diagnostic models. Therefore, we present the largest specialist generative foundation model for chest radiographs to date, with over 1.3B parameters, trained for 1.6T tokens on a curated, heterogeneous dataset comprising 1.2M radiographs and clinical expert-guided metadata. Our model supports controllable radiograph generation and editing across multiple demographic subgroups, acquisition views, and a dozen pathologies. Moreover, we significantly advance the state of the art in radiograph synthesis fidelity, producing images that are indistinguishable from real radiographs to clinical experts.

Foundation models for language and vision are powered by internet-scale data, while structured domains such as tabular prediction are powered by synthetic data. This substitute shifts the challenge from collection to prior design. Such priors already exist for many structured tasks: TabPFN and its successors solve tabular classification with a transformer pretrained on a synthetic Bayesian prior. We make two points. \textbf{First}, reinforcement learning is the conspicuous gap: sampling a synthetic MDP is as feasible as sampling a synthetic tabular dataset, yet no in-context RL work treats prior design as a primary objective. \textbf{Second}, MDPs admit a fixed-size sufficient statistic, independent of the episodes observed and tabular in shape, which makes them directly amenable to the attention-based architectures used for tabular foundation models, with a policy head replacing the supervised target. Together these define the agenda for an RL foundation model. As a proof of concept, we train a Graph Attention Network entirely on synthetic MDPs and show that, with no task-specific tuning, it solves held-out tabular benchmarks in context, both online and offline: online, in far fewer episodes than UCB-VI and tabular Q-learning, and offline, competitively with VI-LCB.

Zero-shot anomaly detection (ZSAD) is attractive for medical imaging because clinical systems must handle heterogeneous acquisition protocols, changing patient populations, and pathologies for which annotated training data may be unavailable. Most existing zero-shot anomaly detection methods are designed for 2D images, and their direct extension to 3D medical volumes is limited by the scarcity of large-scale volumetric foundation models or by the difficulty of utilizing volumetric context. We propose CS3F, a training-free batch-based framework for ZSAD in 3D medical images using 2D foundation models. Each volume is decomposed along multiple anatomical axes and encoded slice-wise by a 2D vision transformer. These are then converted into localized volumetric tokens by pooling neighboring slice features. Anomaly scores are obtained from cross-subject mutual similarity: tokens that lack close analogues in other subjects are assigned higher anomaly scores. To reduce the attenuation of focal lesion signals caused by depth pooling, we introduce a coarse-to-fine tokenization strategy that enables fine-resolution volumetric scoring without exhaustive matching. CS3F is evaluated on brain MRI across metastases, glioma, and stroke, as well as validated on lung CT to test generalizability beyond atlas-aligned brain MRI. The results show that frozen 2D foundation models can support anomaly localization in 3D medical images, and that the benefit of fine tokenization depends strongly on lesion contrast and imaging modality.

Embodied Foundation Models (EFMs) integrate multimodal understanding, future-state reasoning, and executable robot actions. Yet their safety alignment for human-injury prevention remains underexplored, primarily because real-world data of robots harming humans or creating hazardous household situations cannot be safely or ethically collected. To address this challenge, we propose a safety-critical data construction pipeline for human-injury prevention in EFMs.Starting from real DROID observations, our construction pipeline proceeds through scene understanding, hazard-aware image editing, temporal prompt generation, and single-pass rollout synthesis. The temporal prompts specify the expected scene evolution, while Wan2.7 synthesizes realistic robotic rollouts from the edited hazardous states in a single pass. Using this pipeline, we construct ROBOSHACKLES, a 10,000-clip robotic video dataset derived from real DROID observations, spanning two direct-harm and four indirect-harm categories. To ensure dataset quality, we assess task completion and visual quality with automatic metrics, and evaluate six representative EFMs under a refusal-based safety criterion. Results show that all evaluated models produce unsafe actions in the tested safety-critical scenarios, yielding a 100% unsafe action generation rate. ROBOSHACKLES serves as a scalable benchmark and training resource for refusal learning and hazard anticipation before robot action execution.The dataset is publicly available at https://huggingface.co/datasets/YZW00/RoboShackles.

Knowledge graph (KG) foundation models (KGFMs) are zero-shot generalizers: trained once, they can predict links on unseen graphs without retraining. However, understanding when and how they can robustly generalize across KGs is still an open question. In this paper, we shed some light on their generalization mechanisms highlighting how their performance on unseen KGs is not uniform when it comes to partially seen links, which we call half-links. In fact, we show that to predict a test triple $(h,r,t)$ it might suffice in practice to have observed the half-link $(h,r)$ or $(r,t)$ in the inference graph. This yields a taxonomy of four scenarios when combinations of these half-links are observed or not. In a rigorous stratified analysis over these scenarios, we reveal that SoTA KGFMs use seen half links for predictions, while unseen half-links pose different challenges. As such, our finer-grained taxonomy can be a diagnostic protocol for robust KGFM generalization and highlights where novel KGFMs can improve.

Automated pollen identification from microscopy remains a bottleneck in aerobiology, palaeoecology and biodiversity monitoring, because scalable systems must generalise across specimen preparation, scanner settings and geographic origins while retaining palynological interpretability. To address this gap, we present a million-scale multimodal pollen microscopy resource, Pollen AI Atlas, assembled from pure-species whole-slide bright-field images spanning four geographic origins, four scanner settings and 46 taxon labels across 31 botanical families. Seeded by one manually selected exemplar per source slide, token-level mining and filtering produced 1,511,390 released grain detections with 99.6% proposal precision in expert-curated test regions. Each detection was paired with machine-generated grain-level morphological captions from five open-weight vision-language models, guided by expert-verified palynological anchors, yielding structured descriptions of aperture systems, wall ornamentation, shape and size. Among the evaluated models, Gemma4 provided the most controlled primary caption set, combining tight length control, no leakage and the strongest text-retrieval performance. Baseline benchmarks with frozen visual features reached 88.16% top-1 accuracy, while cross-regional retrieval showed that caption-derived text embeddings remained robust when image similarity degraded (mAP@20 0.811 versus 0.262). Released data, annotations, captions, splits, code, and weights provide a benchmark for pollen recognition, cross-regional domain adaptation and domain-specific multimodal microscopy learning.

In this report, we present LOGOS (Language Of Generative Objects in Science), a scientific generative language model that unifies heterogeneous tasks across the natural sciences within a single autoregressive framework based on a shared scientific grammar. It encodes diverse scientific objects and their spatial interactions as token sequences over a common vocabulary. By representing spatial contact and constraint patterns as discrete tokens, the model captures complex structural interactions in a purely sequential manner, without relying on explicit coordinates or geometric neural networks. This unified representation enables a wide range of downstream tasks to be formulated consistently as next-token prediction in the same grammar space, creating strong alignment between continued multi-domain pre-training and downstream objectives. Across diverse tasks, LOGOS consistently matches or outperforms domain-specific baselines, providing preliminary evidence for the feasibility of "one model fits all" in the natural sciences. We train LOGOS models at different scales (1B, 3B, and 8B parameters) and find a consistent positive correlation between model size and performance. This suggests that the future of AI for Science (AI4S) may not lie in building an independent technical stack that is separated from large language models (LLMs). Instead, it may depend on deeply aligning scientific foundation models with LLMs through shared architectures, shared training paradigms, and shared inference infrastructure, so that LLMs can truly become a new entry point for AI4S. We release the model weights and associated resources to facilitate further research.

Semantic search and recommendation of similar documents, such as news and reports about unusual environmental events (e.g., a dead whale washed ashore in Alaska) that contain spatial and temporal information, is a critical task in Geographic Information Retrieval (GIR). This work presents a novel framework that leverages AI foundation models, including Large Language Models (LLMs) and Vision-Language Models (VLMs), to enable effective similarity search and ranking for such event documents. To support this goal, we introduce two new strategies: (1) CAMERA (Context-Aware Multimodal Event Retrieval Algorithm), which fuses textual and visual information to generate richer embeddings than those derived from text alone; and (2) ASTRA (Adaptive Spatial and Temporal Re-ranking Algorithm), which improves similarity ranking by incorporating scale-dependent spatiotemporal relevance alongside semantic similarity. Experimental results, using a dataset from the Local Environmental Observer Network, demonstrate that our VLM-enhanced methods outperform unimodal, LLM-based approaches in similarity ranking effectiveness. By automatically linking relevant event reports, the proposed framework helps both data curators and the general public gain deeper insights into environmental change and its localized impacts. These findings highlight the potential of AI foundation models to advance GIR through multifaceted, intelligent analysis that integrates key geographic concepts: space, time, scale, and semantics.

AlphaEarth Foundations (AEF) unify global remote sensing foundation embeddings through multimodal self-supervised learning, but their pretraining focuses on physical land-surface signals, limiting plug-and-play use in socioeconomic tasks. We integrate seven heterogeneous data streams across 36 Chinese cities over eight years - AEF embeddings, population, nighttime lights, remote sensing indices, points of interest (POIs), urban morphology, and cross-lingual text - and construct CHN-Econ, a socioeconomic benchmark with 16 labels in three categories. We conduct 31 controlled experiments along five axes: fusion architecture, self-supervised objective, text integration, embedding dimensionality, and normalization. Used alone as a linear probe, AEF achieves R2 values of only 0.301 for cross-region and 0.160 for cross-tier evaluation. The five-axis ablated backbone improves these scores to 0.832 and 0.671, respectively, but reveals that low-dimensional semantic streams are consistently suppressed by high-dimensional streams under shared reconstruction. To address this bottleneck, we propose Capacity-Adaptive Reconstruction (CAR), replacing shared reconstruction with per-stream decoders and stream-level losses to mitigate inter-stream capacity competition. CAR further raises cross-region and cross-tier R2 to 0.848 and 0.693, and restores collapsed labels from negative R2 to a stable range. Using CAR, we infer 14.4 million pixels across 36 cities and eight years and release AEF-Econ, including 128d and 64d compressed versions. Self-diagnostics and case studies show that AEF-Econ captures cross-city hierarchies and intra-urban spatial organization under unsupervised settings, providing a socioeconomic remote sensing foundation embedding complementary to AEF physical embeddings.

Foundation models are now used in settings where the prompts they receive can change quickly. Users change, topics change, policies change, and the model may suddenly face a kind of request that was rare in the calibration data. This makes fixed calibration risky. Conformal prediction and conformal risk control give model-agnostic ways to control error, but they work best when the calibration data still look like the future data. This paper develops PromptShift CRC, a drift-aware conformal risk control method for foundation-model outputs under prompt and domain shift. The method embeds prompts and responses, measures how far the current prompt stream has moved from the calibration pool, gives more weight to relevant or recent calibration examples, and updates the risk level online after observed violations. It reports three practical diagnostics: realized risk error, prompt drift, and effective calibration size. We give conditions under which the method controls risk up to terms for distribution mismatch and weighted quantile uncertainty. In a synthetic prompt-shift benchmark, static conformal risk control fails sharply after drift, while PromptShift-CRC gives the best coverage among the adaptive baselines considered. We then evaluate the same calibration layer on public benchmark derived streams for question answering, toxicity, summarization factuality, and long-context hallucination risk

Organ segmentation from PET/CT is critical for quantitative analysis and radiotherapy planning in oncology. To ease the high annotation cost of PET/CT segmentation, semi-supervised learning (SSL) provides a practical and effective solution for developing deep models with limited labeled data. Recent developments in visual foundation models have demonstrated remarkable adaptability with improved efficiency. In this work, we propose a mutual distillation framework that seamlessly exploits both structural and functional foundation models, which act as modality-specific generalists for distilling knowledge from structural CT and metabolic PET imaging. By bridging the gap between the task-specific precision of student models and the segmentation priors of generalist foundation models, we propose \textbf{MuDuo}, a mutual distillation framework that synergistically leverages SAM-Med3D for CT and SegAnyPET for PET to distill their knowledge into a lightweight student network. Our approach eliminates the need for manual prompts while maximizing the utility of unlabeled data for automatic segmentation, achieving state-of-the-art performance on the AutoPET dataset with only 5 labeled cases. Our source code is available at https://github.com/Wu-beining/MuDuo.

Foundation machine learning interatomic potentials (MLIPs) have enabled atomistic simulations across broad regions of chemical and materials space, but many remain computationally expensive and lack explicit electrostatics, limiting their use for systems governed by long-range interactions and electrical response. Previously, we introduced Latent Ewald Summation (LES), which learns latent atomic charges and long-range electrostatics from density functional theory (DFT) energy and force labels alone. Here, we use LES to extract electrostatics that are latent in foundation models: energies and forces predicted by a teacher model are used to train a lightweight LES-augmented student MLIP, with optional fine-tuning on additional DFT data. The resulting models reduce computational cost while providing access to Born effective charge tensors, and infrared spectra. We benchmark student models distilled from a broad set of foundation MLIPs, including UMA, MACE, Orb, eSEN, GemNet-OC, PET, and EquiformerV2-based models, against experimental infrared spectra for liquid water, concentrated hydrochloric acid, and the anatase TiO2(101)-water interface. Across these systems, electrostatic response can be extracted from most foundation MLIPs. The benchmark further shows that the underlying DFT level and dataset used to train the teacher model play a larger role than architecture in determining electrostatic and spectroscopic accuracy. For the TiO2-water interface, fine-tuning with a modest amount of higher-level DFT data improves structural and infrared predictions. LES-based distillation therefore provides a practical route for converting foundation MLIPs into efficient, electrically responsive models, while also testing the physical fidelity encoded in foundation models.

A central challenge in EEG authentication is that models are typically tied to the acquisition settings in which they are trained. In particular, variations in headset hardware, channel layout, and signal duration create heterogeneous recordings that existing models are not designed to handle, causing each new headset or dataset to be treated as a separate model-development problem. This fragmentation limits multi-dataset learning, hinders knowledge transfer, and reduces model reusability. To address this limitation, we present NeuroShield, a reusable foundation model for EEG authentication that learns identity-discriminative embeddings from variable-channel and variable-length EEG recordings through a dual-stage transformer architecture. We pretrain NeuroShield on three public EEG datasets comprising 15{,}762 subjects and 28{,}116 sessions, and evaluate transfer on two unseen downstream datasets. Our evaluations show that, after fine-tuning, NeuroShield reduces equal error rate by 0.44--8.06 percentage points relative to the state of the art. NeuroShield further generalizes to segments longer than those seen during training and operates across channel layouts not encountered during pretraining. These results establish NeuroShield as a reusable and adaptable EEG identity encoder across heterogeneous recording settings. We release NeuroShield as open source to support reproducibility and community adoption.

Wireless foundation models offer a path toward reusable channel state information (CSI) intelligence for sixth-generation (6G) systems. However, existing generic-backbone adaptation and CSI pretraining methods often treat CSI as task tensors rather than propagation-conditioned channel responses, thereby failing to capture the intrinsic time-frequency-spatial geometry of wireless environments. This paper presents a channel-adaptive roadmap toward CSI-native foundation models, proposing a unified framework that aligns pretraining, positional modeling, and attention control with three channel requirements: scale-aware heterogeneous exposure, physical time-frequency-antenna coordinates, and correlation-bounded token interaction. Extensive experiments demonstrate the superiority of the proposed framework across three dimensions: zero-shot generalization, reducing NMSE by more than 4 dB across spatial-temporal-frequency tasks; scale extrapolation, yielding up to a 5.4 dB gain under 8 times unseen antenna scaling; and inference efficiency, accelerating mobility-aware processing by up to 18.8%. A system-level evaluation with Sionna SYS further shows that the proposed framework uses only 7.01% of dense-pilot overhead, reaches -18.64 dB average NMSE, and improves average net spectral efficiency by 36.6% over dense LMMSE and 15.5% over WiFo, indicating that CSI-native representation learning can support pilot-efficient radio access.

We present CacheRL, a system for training small agent foundation models that achieves 92 percent process accuracy on multi-step tool-calling tasks, approaching GPT-5's 94 percent while requiring 100 times less compute. Our approach addresses three challenges in practical agent training: transferring tool-calling knowledge from large models at scale, enabling reinforcement learning without costly live tool execution, and learning robustly from noisy cached environments. CacheRL introduces three key innovations. First, a hybrid thinking trajectory pipeline augments agent trajectories with LLM-generated reasoning traces, producing training examples that teach models not only what tools to call but also why. Second, the CacheAgentLoop eliminates live execution costs through a three-tier fuzzy cache while preserving trajectory fidelity using token-level masking. Third, a cache-tier-aware reward dynamically adjusts answer-quality weights to avoid penalizing models for cache-induced limitations. Through iterative supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO), CacheRL improves Qwen3-4B-Thinking's validation reward from 0.43 to 0.78. On public agentic tool-calling benchmarks, our model achieves competitive performance against frontier models such as GPT-5. Ablation studies show that removing knowledge transfer reduces performance by 41 percent, while cache-aware rewards contribute a 17 percent improvement. Interestingly, reinforcement learning improves training stability but yields limited gains beyond strong supervised fine-tuning, suggesting that data quality and reward design play a more important role than complex optimization methods in building practical small agent models.

Self-supervised foundation models have shown strong promise in medical imaging. However, existing MRI foundation-model studies have primarily emphasized segmentation and dense prediction tasks, while systematic investigation of self-supervised foundation models for MRI-based disease detection remains limited. In this work, we investigate two major self-supervised pretraining paradigms for MRI-based disease detection: reconstruction-based learning via Masked Autoencoders (MAE) and predictive representation learning via Joint Embedding Predictive Architectures (JEPA). We study the role of auxiliary objectives by introducing a novel spectral-domain reconstruction loss for MAE to enhance sensitivity to fine-grained anatomical structure, and by integrating variance--covariance regularization (VCR) within our JEPA framework to encourage decorrelated latent representations. Our models are pretrained on heterogeneous single-contrast MRI volumes in a contrast-agnostic setting, without modality concatenation. Across five downstream disease detection tasks, our results highlight the importance of self-supervised objective design for medical foundation model pretraining, demonstrating that the downstream benefit of each objective is determined by its relevance to the task's structure. Specifically, spectral regularization yields the largest improvements when the downstream discriminative signal is characterized by strong high-frequency anatomical structures, while covariance regularization is most beneficial when discriminative information spans multiple decorrelated feature dimensions. MAE with spectral-domain supervision consistently achieves superior downstream performance for MRI-based disease detection. These findings suggest that self-supervised objectives in medical imaging encode specific biases, and their downstream benefit is fundamentally conditioned on the task's structure.

AI systems are being deployed across medical imaging faster than their failure modes are understood. At this point in time, the failure of greatest clinical concern is hallucination: clinically plausible but factually incorrect outputs, including fabricated anatomical structures, missed findings, incorrect laterality, and invented measurements in generated reports, with direct consequences, for example, for biopsy decisions, staging, and treatment planning. This structured narrative synthesizes peer-reviewed studies, benchmark datasets, and FDA regulatory guidance across five imaging modalities to produce a cross-modality analysis of hallucination taxonomy, etiology, detection, and mitigation. Specifically, we address three questions in this study: (1) how can existing taxonomies be unified across modalities?, (2) how do medical-specialized foundation models hallucinate less than general-purpose ones?, and (3) which mitigation strategies are effective and compatible with FDA lifecycle oversight? We note that three taxonomic frameworks together cover the imaging pipeline in a way no single framework does alone. We also highlight that general-purpose foundation models outperform medical-specialized models on hallucination-specific benchmarks, indicating that narrow domain fine-tuning can introduce overfitting-induced confabulation. At the same time, the oversight of radiologists remains essential; for instance, a very high percentage of of AI-generated flags required expert correction before clinical use. Physics-informed architectural constraints, Chain-of-Thought prompting, and human-in-the-loop safeguards each address different failure modes and is effective when combined. All findings are mapped to the FDA's Total Product Lifecycle and Predetermined Change Control Plan frameworks, which treat hallucination management as a lifecycle obligation rather than a pre-deployment checklist.

Environmental scientists spend disproportionate effort on data wrangling rather than analysis. New AI agents can be a helpful tool, but no benchmark exists to evaluate AI agents that automate environmental geospatial workflows through structured tool calling against real APIs. We introduce the GeoNatureAgent Benchmark, the first benchmark for environmental analysis agents that operate via structured tool calls to a production-style geospatial API. The benchmark comprises 93 tasks across 18 categories. Tasks are evaluated against an open, self-hostable geospatial API that serves three environmental indicators across Spain and Portugal via sixteen tools. We evaluate nine frontier and open-weight LLMs, reporting capability and per-case cost as orthogonal axes. Results manifest that (1) Claude Sonnet 4 achieves the highest capability at 60.8% +/- 0.8%, followed closely by DeepSeek V3.2 at 56.3% +/- 3.1%, while no other model exceeds 51%; (2) the cost-accuracy Pareto frontier is occupied mostly by open-weight models, with DeepSeek V3.2 offering 93% of Claude's capability at 11.6x lower cost; and (3) structured tool calling against a real API provides a more discriminative measure of real-world agent capability, with mean accuracies 25-35 percentage points below those reported on general-purpose GIS benchmarks.

Recent progress in foundation models has shifted toward agentic behavior involving multi-step reasoning and tool use. However, open-source efforts largely focus on text-dominant settings, leaving long-horizon multimodal tasks underexplored. This gap is evident in video tasks requiring sustained temporal understanding and iterative interaction. We present InternVideo3, a framework enhancing these capabilities via Multimodal Contextual Reasoning (MCR). MCR treats understanding as a closed-loop process over a shared, evolving context containing observations, instructions, reasoning, tool actions, and memory. This frames long-video understanding as evidence accumulation and verification. To ensure efficiency, we introduce Multimodal Multi-head Latent Attention (M^2LA), a token-preserving reparameterization compressing KV-cache states while retaining the full token stream. Our staged training includes continued pretraining, short-to-long supervised fine-tuning, rule-based reinforcement learning, and on-policy distillation. Experiments show InternVideo3 achieves strong performance on benchmarks like Video-MME, MLVU, and EgoSchema. We further instantiate the model as a video agent with retrieval tools, demonstrating robust evidence-grounded behavior. Our results suggest that efficient context handling and closed-loop reasoning are vital for adapting open multimodal models toward long-horizon visually grounded agency.

This paper presents a novel data-free and training-free compression approach for speech foundation models using channelwise clustering via k-means. More fine-grained, mixed sparsity pruning by layer-level varying number of parameter clusters is also explored. Experiments conducted on the LibriSpeech dataset suggest that when operating with pruning sparsity of 50% on HuBERT-large, consistent WER reductions of 27.73%/18.61% absolute (34.37%/21.91% relative) over the magnitude-based pruning were obtained on the test-clean and test-other subsets before fine-tuning and 0.19%/0.79% absolute (3.36%/4.62% relative) after fine-tuning with only 3 epochs. Similar WER reductions of 2.86%/5.02% absolute (59.21%/55.29% relative) were observed against magnitudebased pruning on Whisper-large-v3 at 10% sparsity, all with no significant WER increase relative to the uncompressed baseline.

We study controlled post-training refusal suppression in routed MoE and hybrid-MoE foundation models, aiming to increase non-refusal target-response behavior while preserving general capability under a compact intervention footprint. Existing broad direction-based edits can perturb general-purpose computation, whereas support-only expert edits often lack sufficient capacity to correct heterogeneous refusal representations. To address this limitation, we introduce Localized Multidirectional Correction (LoMC), a support-gated intervention framework that follows a support-then-correction execution order: it first identifies a compact edit support, then aggregates prototype correction directions into layer-wise correction directions, and finally applies rank-one layer-wise correction only within the selected support. By using the edit support as a structural gating constraint, LoMC increases correction capacity without expanding the intervention scope. Experiments on text-only and multimodal safety benchmarks across four routed backbones show that LoMC substantially improves non-refusal target-response behavior while maintaining general capability under a compact intervention footprint.

The deployment of Time-Series Foundation Models (TSFMs) in physical sciences is hindered by a critical trade-off: while these models encode rich, universal temporal dynamics, they suffer from severe distributional misalignment when applied zero-shot to specific scientific domains, and their computational cost prohibits deployment in edge-computing sensor networks. We address a fundamental challenge: How can we extract latent structural knowledge from misaligned foundation models (FM) to train lightweight, specialized forecasters? We propose Gated Uncertainty-Aware Routing for Distillation (Guard), a novel framework that reframes multiteacher distillation as an instance-wise decision process with two adaptive mechanisms: (1) a Contextual Router that dynamically selects the most relevant teacher based on local input statistics, exploiting complementarity across diverse foundation models; and (2) an Uncertainty-Gated Temperature mechanism that acts as a "circuit-breaker," automatically attenuating distillation strength when teacher confidence diverges from domain reality. We evaluate our proposed lightweight framework on four climate-critical domains: meteorology, ecosystem carbon flux, soil moisture, and energy grids. Our method significantly reduces RMSE relative to a fixed-weight multi-teacher distillation baseline, successfully distilling knowledge from pretrained FMs (teachers) even when they exhibit suboptimal zero-shot accuracy due to distribution shift between the original and target data domains. We demonstrate that these domain-misaligned teachers can still serve as critical correctives, outperforming the globally superior FMs on 28.5% of the hardest instances. Ultimately, this enables high-precision scientific forecasting suitable for resource-constrained edge deployment. Code is available at https://github.com/RupasreeDey/GUARD-KDD2026.

Generative AI emulators are increasingly used in scientific domains where we already have strong theory, benchmarks, and physical intuition. This raises a central evaluation and interpretability question: when a foundation-style model can reproduce known continuum dynamics, what internal mechanism supports that behavior, is the internal behaviour consistent with known physics, and how does it relate to where the emulator succeeds or fails? We investigate a cross-domain foundation model for continuum dynamics, Walrus by Polymathic, using mechanistic interpretability guided by physical principles. We apply a sparse autoencoder (SAE) to probe a selected layer, and address the practical challenge of triaging a large feature set (over 20,000) using enstrophy as a physically grounded metric. As a deliberately simple testbed, we focus on shear flow and compare feature recruitment across multiple shear-flow setups, i.e. parameter values in the numerical simulation. Across setups we find evidence of piecewise consistency, with subsets of features recurring in similar roles, but this structure is intermittent and does not map cleanly onto standard physical decompositions. In parallel, direct comparisons between numerical simulation and the emulator reveal systematic output-level discrepancies, including regimes where energy/structures become too diffuse or too localized. We connect parts of these discrepancies to changes in specific SAE feature usage. Our work highlights open questions for scientific foundation models: how to robustly prioritize mechanistically meaningful features, how to separate stable structure from analysis artifacts (including single-layer and SAE limitations), and how to use established benchmarks to decide when "different" internal representations are genuinely informative rather than merely effective.

Transfer learning from large-scale RGB foundation models to infrared (IR) imagery through knowledge distillation (KD) remains challenging due to fundamental differences in image formation physics. We investigate the spectral structure of the RGB--IR modality gap and observe that feature divergence is not uniform across spatial frequencies: low-frequency components (shape, layout) show greater cross-modal alignment than high-frequency components (texture, fine edges), which reflect modality-specific characteristics. Based on this analysis, we propose FreqKD, a frequency-decoupled distillation framework that applies asymmetric supervision adapted to each band's cross-modal consistency. The method employs strict mean squared error (MSE) on the low-frequency band to preserve shared structural information and a relaxed log-MSE loss (weighted at 0.1) on the high-frequency band to provide edge guidance while tolerating texture differences. Spectral divergence analysis on 500 paired samples shows that high-frequency divergence exceeds low-frequency divergence by a factor of 2.4x on average across all analysed transformer layers. On KAIST multispectral pedestrian detection, FreqKD achieves 64.1 mAP50, improving 2.4 points over the DINOv2 baseline. The learned representation transfers across datasets (FLIR ADAS, +2.1 mAP50), tasks (MFNet segmentation, +1.85 mean intersection-over-union), and architectures (ResNet-50, +1.0 mAP50). Code is available at: https://anonymous.4open.science/r/freq_decoupled_kd-5E5A

We introduce Embodied-R1.5, a unified Embodied Foundation Model (EFM) that integrates comprehensive embodied reasoning capabilities, spanning embodied cognition, task planning, correction, and pointing, within a single architecture toward general physical intelligence. Leveraging three automated data construction pipelines to significantly expand the data coverage of critical capabilities, we build a large-scale data system of over 15B tokens, and design a multi-task balanced RL recipe to alleviate heterogeneous task conflicts. We further introduce a Planner-Grounder-Corrector (PGC) closed-loop framework that enables a single model to autonomously execute and self-correct over long-horizon tasks. With only 8B parameters, Embodied-R1.5 achieves SOTA on 16 out of 24 embodied VLM benchmarks, surpassing leading models like Gemini-Robotics-ER-1.5 and GPT-5.4. Benefiting from the internalized embodied capabilities, Embodied-R1.5 can be fine-tuned into a VLA with only a small amount of data, outperforming leading VLA models like $π_{0.5}$ across 4 popular manipulation benchmark suites. We further conduct extensive zero-shot real-robot experiments, validating performance in instruction following, affordance grounding, articulated object manipulation, and long-horizon complex tasks, demonstrating strong generalization to the physical world. We open-source model weights, datasets, training code, and EmbodiedEvalKit, an evaluation framework tailored for embodied tasks, to facilitate future research in EFMs.

Foundation Models (FMs) have recently redefined the state-of-the-art in histopathology by providing robust representations for whole-slide image (WSI) analysis. However, selecting the optimal foundation model (FM) for a specific clinical cohort currently requires multiple preprocessing steps, followed by computationally expensive feature extraction and the training of a Multiple Instance Learning (MIL) aggregator for every model. In this work, we investigate whether efficient tile-level linear probing can serve as a reliable proxy for slide-level performance, reducing the need to run full slide-level pipelines for every candidate encoder. We benchmark 19 state-of-the-art FMs on 42 slide-level and 16 tile-level tasks, comparing tile probing metrics against slide-level outcomes using ABMIL and Mean Pooling aggregations. We observe a high correlation between tile and slide performance across varying task difficulties, indicating that encoder representation quality is the primary determinant of WSI success. Sensitivity analyses show that transferability is stable across models and is more influenced by cohort sizes and numbers of tiles per slide than by average task difficulty. We also measure the agreement in best performing models between tile and slide-level tasks, showing tile benchmarks reliably shortlist strong candidates. Overall, our study indicates that tile-level benchmarking provides an efficient and practical first step for narrowing down candidate models, while slide-level evaluation remains essential for final validation on clinical tasks.

Though wireless foundation models (WFMs) have shown strong potential in learning universal channel representations, their adaptation to various downstream tasks remains constrained by existing paradigms. Fine-tuning strategies introduces substantial computational and storage overhead, while frozen feature extraction leads to sub-optimal performance across diverse downstream tasks. To address this issue, we propose a unified adaptive feature composition framework for multitask generalization in WFMs, where the key component is the Routing Adapter for Feature Composition (RAFC). Instead of extracting only the final-layer output, this router treats the hidden states from different Transformer depths as a reusable pool of multi-level hidden features, and employs a lightweight task-driven feature composition network to generate layer-wise aggregation weights, then adaptively combine hierarchical representations through weighted summation. This design enables each downstream task to access suitable mixture of low-, mid-, and high-level wireless features without modifying the pretrained backbone. Extensive experiments on four representative wireless tasks demonstrate that RAFC consistently outperforms conventional adaptation baselines while introducing fewer than 50K additional parameters. Moreover, the learned routing weights provide interpretable evidence of task-specific layer preferences, making the proposed framework a low-complexity, scalable, and explainable interface for adapting WFMs to diverse downstream scenarios.

Foundation models (FMs) are increasingly used as backbones for downstream tasks across language, vision, time-series, and multimodal applications. Yet existing model-serving systems deploy each customized task as an independent model instance, thereby replicating heavyweight backbones, wasting accelerator memory, and losing opportunities to amortize batching and loading costs. This paper presents FMplex, a serving system that treats FM backbones as a virtualization substrate for deployment sharing. FMplex presents each task with a virtual foundation model (vFM), a logically private FM instance backed by a shared physical FM. This abstraction lets independently customized tasks share a backbone while preserving task-specific extensions, independent lifecycles, and task-level isolation. In addition, we propose a batch-aware fair-queueing scheduler that combines weighted task-level sharing with inter- and intra-task batching across colocated tasks. We implement a FMplex-based serving stack spanning task construction, sharing-aware deployment, and runtime execution. Across 7 FM backbones (16 variants) and 92 downstream tasks, FMplex reduces latency by up to 80% over spatial partitioning and 33.3% over best-effort co-location, while hosting up to 6x more tasks at cluster scale.

Self-supervised data curation provides a pathway to scaling and improving the generalization capabilities of machine learning models. By leveraging self-supervised learning (SSL) for data curation, the demand for massive training datasets required by foundation models can be effectively met. SSL greatly alleviates the costs associated with annotation and manual dataset curation while minimizing the need for human oversight. However, the integrity of SSL-curated datasets must be rigorously checked, as reliance on anonymous and unvetted external sources can substantially increase the risk of data poisoning. In this paper, we propose a Poisoned Data Detector (PDD), an active defense mechanism designed to ensure the integrity of SSL-curated datasets prior to foundation model training. PDDs are designed using a combination of the pretrained ImageBind model and traditional classifiers, including Random Forest (RF), k-Nearest Neighbors (KNN), Naive Bayes (NB), and Support Vector Machines (SVM). We rigorously evaluated PDDs using 176,200 images from three diverse datasets and three different adversarial attacks encompassing both in-distribution and out-of-distribution scenarios. Notably, SVM-PDD achieves superior performance for both in-distribution (Set3-Set5) and out-of-distribution (TrueFace and 140K RealFace) datasets. Our design demonstrates strong scalability and enables the rapid integration of new adversarial attack detectors through an ensemble approach.

Tensor networks provide efficient representations for compressing large neural networks. By carefully designing shapes and topologies, they can significantly reduce memory and computational costs. However, identifying implicit low-rank structures in large foundation models remains challenging due to their enormous scale and un-structured weight distributions. We propose an adaptive tensorization method that discovers inherent low-rank structure in a target tensor by index ordering. Experiments on weight and KV-cache compression demonstrate improved reconstruction quality compared to baselines.

With the evolution of large foundation models (LFMs), data-driven autonomous driving has made significant strides. However, existing paradigms still face severe challenges in complex interaction and long-tail scenarios due to distribution shift and causal confusion. These limitations often result in a lack of human-level decision-making flexibility and safety in extreme conditions. To overcome this limitation, this paper proposes a Human-in-the-Loop Online Behavior Cloning frame work (HiL-OBC) for autonomous driving, which aims to deeply integrate the cross-modal perceptual capabilities of LFMs with the high-level driving intelligence of human experts. Specifically, HiL-OBC deployment is executed through three critical phases: policy initialization with human intervention, latent behavioral modeling with Bayesian policy adaptation, and online deploy ment and updates. Furthermore, we design a Multi-modal Online Behavior Cloning (MOBC) model, which optimizes the base driving policy online through a lightweight network architecture, a takeover trigger mechanism, and a multi-variant loss function, thereby enhancing the system's decision-making robustness in complex environments. We evaluated the HiL-OBC on the LangAuto-Human CARLA benchmark. Experimental results demonstrate that the driving policies optimized via the human-in-the-loop mechanism achieve substantial performance gains: the DS of StructNav, LFG, and LMDrive increased by 47.25%, 31.59%, and 32.12%, respectively, with a simultaneous of various experimental settings and key components highlights the advantages of human-in-the-loop learning in improving decision-making robustness and overall driving performance.

Displacement time series from Global Navigation Satellite Systems (GNSS) are essential for a wide range of applications, including monitoring tectonic crustal deformations and investigating the different stages of the earthquake cycle. Machine learning methods have proven promising for GNSS applications; however, most remain fully supervised. This creates a bottleneck as labeled data are scarce, even though large amounts of unlabeled GNSS data are freely available. We present GNSS-FM, a self-supervised foundation model for daily GNSS time series. The model uses a dual-stream input combining displacement and velocity-like increments, and is pretrained using a masked latent prediction objective with vector-quantized targets adapted from wav2vec 2.0, with several modifications for geodetic data. Pretrained on data from over 17,000 globally distributed GNSS stations, an analysis of the learned codebook suggests that the representations capture the main signal types in GNSS displacement data, including seismic offsets, tectonic drift, and seasonal patterns. The foundation model is later fine-tuned on two downstream tasks, namely 90-day displacement forecasting and seismic step localization, where it outperforms strong task-specific baselines in both cases. These results show that self-supervised pretraining is a promising approach for GNSS time series analysis.

Progress in AI-driven crystal materials science has so far been carried by narrow architectures purpose-built for individual tasks -- graph neural networks for property prediction, diffusion and flow-matching models for crystal generation -- each excelling within its niche yet unable to act as a shared backbone across the full spectrum of materials problems. Generative large language models offer a fundamentally different paradigm, in which structural representation, quantitative prediction, and structure-activity reasoning can be unified within one model, but the materials community has yet to see this paradigm realized at a level competitive with established narrow specialists. Here we present MatMind, a generative foundation model purpose-built for crystal materials science under this paradigm, developed through the coordinated activation of structure-activity knowledge and physics-informed feedback within a progressive training framework -- combining structure-activity knowledge injection, a dual-head architecture that jointly trains language reasoning and numerical regression in a shared representation space, and multi-objective physics-informed reinforcement learning over stability, novelty, and structural diversity. Across three task families, MatMind attains the lowest mean absolute error on energy above hull, bulk modulus, and band gap -- surpassing graph neural network predictors purpose-built for these tasks -- reaches an S.U.N. rate of 65.3% on unconditional crystal generation, and achieves a comparable multiplicative improvement on magnetization-density-conditioned generation, where only 21 positive samples exist within over 600000 training entries. By matching or surpassing narrow specialists on their own ground while operating within a single unified model, MatMind shows that the LLM-based paradigm can serve as a viable backbone for crystal materials science going forward.

Model merging combines several independently fine-tuned experts into a single multi-task model without any training data, reducing the storage, serving, and decentralized-development costs of large foundation models. State-of-the-art merging methods formulate merging as a layer-wise quadratic interference minimization problem. Although this problem admits an exact closed-form pseudoinverse solution, that solution underperforms hundreds of iterations of gradient descent in practice. The iterative loop dominates the cost of the pipeline, yet its effectiveness has remained unexplained. We revisit this regime and show that the iterative solver does not primarily act as an optimizer; rather, it serves as an implicit spectral regularizer for an ill-posed normal equation, where small-eigenvalue directions of the per-layer interference operator amplify proxy noise. Building on this finding, we formalize multi-task model merging as a noisy linear inverse problem and propose a spectral filtering estimator parameterized by a per-direction filter. We instantiate this estimator with SWUDI, a closed-form method that combines a soft exponential filter, which matches the gradient-flow trajectory of iterative descent, with a hard top-K truncation that suppresses noise-amplifying small-eigenvalue directions. Furthermore, we propose SWUDI-A, an adaptive variant that replaces the global rank hyperparameter with per-layer rank rules, further improving robustness across architectures. Both variants share a single symmetric eigendecomposition per linear layer and require no training data or optimizer state. Across four general benchmarks and a multimodal merging benchmark spanning VQA, Geometry, Chart, OCR, Grounding, and modality merging, our proposed spectral solvers match or outperform state-of-the-art merging methods. Crucially, they reduce wall-clock time by 28-72x and peak GPU memory by up to 50%.

Foundation model agents are increasingly deployed for real-world decision-making, but suffer from the sim-to-real gap. While robotics and classical control have mature frameworks to address this gap, the foundation model community is treating agent robustness as an entirely novel phenomenon. Our paper proposes formalizing the foundation model agent evaluation and training gap as a classical sim-to-real problem structured entirely around the four elements of a Markov Decision Process, including Observation, Action, Transition, and Reward. In this paper, we set a comprehensive research agenda that translates classical discrepancies into the foundation model domain and advocates for adopting established solutions like domain randomization. We provide concrete examples, such as a multilingual tool calling to demonstrate how severe observation space gaps lead to operationally invalid actions despite correct semantic intent. Ultimately, this agenda aims to drive a paradigm shift, yielding a unified vocabulary and standardized stress test benchmarks to foster a new generation of highly trustworthy agents for reliable real-world applications.

Wireless foundation models have emerged as a promising alternative to building separate models for each wireless task. However, existing approaches rely on masked input reconstruction, which can bias representations toward low-level signal details. In this paper, we propose LatentWave, a wireless foundation model pretrained using a Joint-Embedding Predictive Architecture (JEPA) on diverse wireless spectrograms and channel state information (CSI). By predicting masked regions in latent space, LatentWave learns representations that are more transferable out of the box across diverse downstream tasks. The proposed architecture employs per-channel patch embeddings with stochastic channel sampling during pretraining, allowing it to process variable antenna counts and improving usability across heterogeneous wireless configurations. We evaluate LatentWave on four downstream tasks: RF signal classification, 5G NR positioning, beam prediction, and LoS/NLoS classification, comparing against a masked-modeling baseline (WavesFM) pretrained on the same data. Additionally, we show that the masking geometry introduces a task-dependent inductive bias: frequency masking strongly favors channel-related tasks such as positioning and beam prediction, while region masking better preserves discriminability for signal classification.

Federated fine-tuning of foundation models using Low-Rank Adaptation (LoRA) offers a communication efficient solution for distributed learning. However, existing federated LoRA methods suffer from two fundamental limitations: (1) structural aggregation bias, where independently averaging low rank factors fails to approximate the true combined update, and (2) client side initialization lag, as clients repeatedly reinitialize LoRA parameters across communication rounds, slowing convergence. We propose HyperLoRA, a unified framework that addresses both issues through amortized federated adaptation through hypernetwork-driven LoRA generation and product space aggregation. Instead of iterative per-client optimization, HyperLoRA employs a learned generator that maps client distribution signatures to LoRA initializations, effectively amortizing per client adaptation. On the server side, we introduce a learned aggregation module that directly synthesizes updates in the low-rank product space, eliminating the inconsistencies of factor-wise averaging. A lightweight residual correction module further improves stability under heterogenous (non-IID) client distributions.By replacing iterative optimization and heuristic averaging with learned operators, HyperLoRA jointly enables efficient personalization, unbiased aggregation, and faster convergence. Experiments on federated vision and vision-language benchmarks show that HyperLoRA achieves improved convergence speed, greater robustness to distribution shift, and stronger personalization performance compared to prior federated LoRA methods.

Entity alignment (EA) aims to identify equivalent entities across heterogeneous knowledge graphs (KGs) and is a key component of knowledge fusion and cross-KG reasoning. The recent EA foundation model demonstrates that alignment knowledge, once pretrained, can be directly applied to diverse previously unseen KG pairs. However, it still underuses structural context in two places: cross-KG interaction is weak during encoding, and final candidate ranking still relies too heavily on coarse similarity. We address these limitations with ContextEA, an enhanced encoder-decoder framework for transferable EA. On the encoder side, we introduce a cross-KG interaction encoder that unifies the two KGs with anchor bridges and performs earlier relation-aware cross-graph propagation. On the decoder side, we introduce a structural calibration decoder that calibrates alignment scores with entity-level, neighborhood-level, relation-level, and anchor-aware structural evidence. This design strengthens both structural context construction and structural context exploitation while remaining lightweight. Experiments on 29 EA datasets in OpenEA, SRPRS, and DBP show consistent gains over strong transferable baselines. Notably, the pretrained ContextEA already surpasses the finetuned baselines on all three benchmark groups, demonstrating substantially stronger transfer to unseen KGs. These results suggest that explicitly harnessing structural context is an effective direction for improving EA foundation models.

Reliable crack assessment requires not only accurate pixel-level masks but also connected crack geometry and confidence estimates that remain stable under domain shift. However, existing segmentation models can achieve high overlap scores while fragmenting cracks, missing fine branches, and providing no calibrated uncertainty. To address this gap, this paper proposes CrackGeoFM, a multi-task framework that combines a frozen visual foundation backbone with crack-specific adaptation for mask prediction, skeleton reconstruction, and uncertainty estimation. The framework integrates a Frequency-Guided Crack Enhancement Module (FCEM) to enhance high-frequency crack cues, a Crack-Domain Feature Adaptation Module (CFAM) to adapt frozen backbone features to crack-domain patterns, and a Structure-Aware Multi-Task Decoder (SMTD) to jointly decode masks, skeletons, and uncertainty. Across 20 crack datasets, CrackGeoFM achieves state-of-the-art segmentation, improved topology preservation, calibrated uncertainty, and effective few-shot adaptation with only five labeled images. These results support reliable, generalizable, and engineering-oriented crack analysis for infrastructure assessment.

Diffusion MRI (dMRI) tractography is the only noninvasive approach for mapping white-matter pathways in the living human brain. It represents each brain as a tractogram: a large, unordered set of three-dimensional streamlines that includes information about both local streamline geometry and whole-brain anatomical organization. This structure makes tractograms a natural but challenging target for representation learning. Existing methods treat streamline classification and subject-level prediction as separate problems: streamline classifiers focus on geometric patterns, whereas subject-level prediction often depends on hand-crafted features. As a result, current methods do not learn reusable representations that connect streamline anatomy with whole-brain inter-subject variation. Here we introduce TractFM, a tractogram foundation model that learns reusable representations directly from whole-brain streamline sets. TractFM combines a local streamline encoder with a permutation-equivariant tractogram encoder, allowing all streamlines from a subject to be contextualized jointly in a single forward pass. Pretraining on dense anatomical tract parcellation, i.e., assigning anatomical labels to individual streamlines, yields two complementary representations: contextualized streamline-level embeddings for tract parcellation and compact subject-level descriptors for downstream prediction of subject phenotypes. Across three tractography algorithms and five dMRI datasets, TractFM transfers to both streamline-level and subject-level tasks. Its frozen representations achieve accurate tract parcellation and predict age and sex across independent datasets. These results show that whole-brain geometric context, learned once, can generalize across tractography pipelines, datasets, and prediction tasks.

Foundation models are increasingly trained on sequences of user actions in recommendation, payments, fraud, and commerce, but these models still lack the kind of compute calibration that scaling laws provide for language models. We study a common two-part behavioral-model architecture: a feature-based event embedder maps each multi-modal item to a vector, and a decoder-only transformer predicts the next event from the resulting sequence. Across roughly 600 runs on real interaction data, spanning $10^{15}$-$10^{19}$ training FLOPs, we jointly vary four deployment-relevant axes: the two-part parameter split, critical batch size, model/data allocation, and the number of sampled negatives used after freezing the embedder. A small embedder ($s^{\star}\!\approx\!2\%$ of parameters) is compute-optimal at every budget we test because embedder parameters are both more expensive per step and exposed to far more repeated items than contextualizer parameters. Compute-optimal training is data-heavy relative to text at low compute, but its $D/N$ ratio moves toward the Chinchilla heuristic as compute increases. The sampled training objective and deployed ranking metrics disagree in ways that themselves scale: critical batch size, optimal negative count after freezing, and the agreement between loss and ranking quality all shift with compute and with the chosen evaluation metric. For negative sampling, larger budgets increasingly prefer more negatives; by $10^{19}$ FLOPs the active constraint is candidate-axis memory rather than FLOPs. In behavioral foundation models, the evaluation metric is therefore part of the scaling law: changing it can change the compute-optimal recipe.

Tabular foundation models (TFMs) have made rapid progress in standard classification and regression, but time-to-event survival prediction tasks have remained largely untouched. Unlike in standard regression tasks, survival prediction models must account for censored data. Standard TFMs cannot handle natively censored data, leading to biased and inaccurate predictions, making them unsuitable for real-world applications. To overcome this fundamental limitation, we propose \texttt{SurvPFN}, a prior-data fitted network (PFN), for survival prediction tasks. We pretrain \texttt{SurvPFN} on millions of synthetic survival prediction tasks to learn survival via distributional regression that accounts for censored data. \texttt{SurvPFN} works by (1) generating data with Weibull event times and a non-informative censoring mechanism; (2) integrating a censored event indicator; and (3) minimizing a censored negative log-likelihood. On SurvSet, a collection of real-world survival tasks, \texttt{SurvPFN} is highly competitive with classical and deep survival baselines without per-dataset fitting, a survival-specific architecture, or feature engineering. We show that survival can be treated as a continuous-time distributional regression problem with censored loss, unlocking the power of PFNs for time-to-event predictions.

Genomic foundation models increasingly adopt large language model architectures, yet almost universally rely on fixed tokenization schemes such as $k$-mers, BPE, or single nucleotides, which impose arbitrary sequence boundaries that may obscure biologically relevant structure. We present LDARNet, a 120M-parameter hierarchical genomic foundation model that adapts H-Net-style dynamic chunking from autoregressive generation to masked language modeling, combining BiMamba-2 state-space layers with local attention, bidirectional routing, and a ratio-based regularizer to induce adaptive token boundaries without supervision. Fine-tuned on 27 tasks from the Nucleotide Transformer and Genomic Benchmarks suites, LDARNet achieves 11/18 wins among compact models ($<$300M parameters) and state-of-the-art results on 5 histone modification tasks, outperforming models up to 20$\times$ larger. A FLOPs-matched controlled experiment isolates learned routing as the source of these gains: learned boundaries beat fixed-grid boundaries by up to 14 percentage points on histone tasks at identical compute. Nucleotide-resolution analysis further shows that the learned boundaries align with canonical promoter motifs and splice junctions without supervision, providing a biological interpretation for adaptive tokenization in genomic foundation models.

Large language models (LLMs) have emerged as powerful foundation models with strong reasoning capabilities across domains. Beyond reactive text generation, agentic LLMs enable autonomous workflow execution through modular task decomposition and coordinated tool use. In structural engineering, recent efforts have developed agentic LLMs for automated analysis of plane frames. However, their extension to 3D frames remains underexplored due to challenges in irregular geometric representation, topological consistency, and long-horizon reasoning. This paper proposes an agentic LLM framework for automated structural analysis of 3D frames from natural language inputs. Irregular 3D frames are represented by projection onto a 2D plan, where orthogonal gridlines define spatial coordinates and a matrix of number of stories encodes vertical extrusion of each grid cell. Building on this representation, the framework establishes a multi-agent pipeline: a problem analysis agent parses input into structured JSON; a floor decomposition agent derives the spatial layout of each floor; the 3D geometry is assembled by node, girder, slab, and column agents; support and load agents assign boundary and loading conditions, and code translation agents generate executable SAP2000 script. Evaluated on ten representative 3D frames, the proposed framework achieves an average accuracy of 90% across repeated trials, demonstrating consistent and reliable performance.

Foundation models have sparked a revolution via a pretraining-adaptation paradigm, with recent efforts extending this success to graphs. Unlike other modalities, graphs contain rich structural patterns, yet their structural transferability remains poorly understood. Prior studies consider common substructures in the discrete realm, and we are motivated by a fundamental question: Are common substructures transferable? The underlying theory is largely underexplored. In this work, we shift toward learning transferable structures through the lens of functional behavior. Theoretically, we connect transferable substructures to intrinsic geometry of the representation space. However, characterizing such intrinsic geometry has rarely been touched. Grounded in Riemannian geometry, we develop a graph intrinsic geometry learning framework called Neural Vector Bundle, which enables parsing intrinsic geometry with local coordinates. Building on this, we design GAUGE, a pretrainable neural architecture that constructs the vector bundle, flattening geometrically compatible local coordinates, and a new Dirichlet loss, which also measures the transfer effort. We empirically validate its superior expressiveness in challenging tasks including zero-shot link prediction and graph isomorphism.

Affordance understanding bridges visual perception and physical action, serving as an explainable interface for robot manipulation in open and unstructured real-world environments. Yet, building an affordance foundation model that not only understands where and how the interaction should happen, but also generalizes across diverse environments, objects, and tasks, remains a long-standing research challenge. Existing methods typically address only part of this challenge, either localizing task-relevant regions without specifying executable motion, or predicting motion but with limited scalability. In this paper, we present ourmodel, a step towards an affordance foundation model for functionality understanding. From a single RGB-D observation and a language task description, ourmodel predicts a task-conditional functional mask (where to interact) and a 3D post-contact motion curve (how to interact). To support open-world generalization, we build a large-scale standardized data pipeline that converts heterogeneous robot, human, simulation, and real-world scan data into a shared affordance schema with language, masks, and object-centric 3D motion labels. We evaluate ourmodel from three aspects: for affordance segmentation, ourmodel outperforms all baselines by a large margin across 8 test sets from 4 benchmarks, improving mean gIoU/cIoU by +23.9/+26.3; for contact-point prediction, it predicts substantially more accurate points, with a 12.7--61.3% hit-rate gain over the best baseline; and for 3D motion, it achieves the best performance on all three test sets. ourmodel can be deployed for real-world robot manipulation without finetuning for robot embodiment or using task-specific heuristics, demonstrating the ability to adapt to open-world affordance tasks. Project page: https://www.zhaoningwang.com/AFUN

Frontier model evaluations are shifting from foundational capabilities (e.g., instruction following and reasoning) toward compositional, agentic ones, but Korean agentic benchmarks remain scarce. We introduce K-BrowseComp, a web-browsing agent benchmark grounded in Korean contexts, consisting of 400 problems. The 300-problem K-BrowseComp-Verified subset is manually constructed and validated by native Korean speakers. On this subset, frontier LLMs, including GPT-5.5, DeepSeek-V4-Pro, and GLM-5.1, reach only 30.00--45.67%, a substantial drop from BrowseComp, while Korean LLMs released through Korea's Proprietary AI Foundation Model program obtain only 0.00--10.33%. We further construct a 100-problem synthetic split using hard few-shot exemplars and failure-mode-targeted generation to exploit the asymmetry between solving and creating web browsing problems. On the adversarially filtered synthetic diagnostic split, the strongest model reaches only 26.00%, and we report this split separately as a targeted stress test. We publicly release our data and code.

Wireless localization is a fundamental capability of sixth-generation (6G) networks. Conventional model-based methods require accurate modeling of the propagation environment and degrade in complex multipath and non-line-of-sight scenarios, while learning-based methods couple model parameters tightly to the training scene, requiring costly retraining whenever the base station (BS) configuration or propagation environment changes. In this paper, we propose RA-LWLM, a retrieval-augmented in-context localization framework that achieves training-free cross-scene adaptation by externalizing scene-specific information into a per-scene fingerprint database rather than encoding it in model weights. The framework consists of three components: a frozen wireless foundation model (FM) encoder that maps raw channel state information into a scene-agnostic representation; a retrieval module that selects the most informative references from the per-scene database via similarity search in the representation space; and a transformer-based in-context learning (ICL) module that fuses the query with the retrieved references to predict the user equipment (UE) position. To accommodate varying retrieval quality and propagation complexity across queries, the ICL module adopts a mixture-of-experts design in which experts specialize in different context sizes and are softly combined by a learnable selector. Extensive ray-tracing-based experiments across heterogeneous scenes with diverse BS configurations show that RA-LWLM achieves nearly identical accuracy on seen and unseen scenes without any per-scene retraining, substantially outperforming end-to-end and FM-based baselines. These results validate the proposed retrieval-augmented in-context paradigm as a scalable solution for cross-scene localization in 6G networks.

Whether physics foundation models can be usefully deployed on laboratory experiments remains an open question for scientific machine learning (ML). We test this question on the Rayleigh-Taylor instability (RTI), a ubiquitous and demanding fluid instability seen from tabletop flows to supernova explosions, in which small perturbations at a density interface grow into chaotic, multiscale mixing as a lighter fluid accelerates into a heavier one. Standard ML models struggle with RTI, and despite over a century of theoretical, numerical, and experimental work, it carries an unresolved discrepancy between simulation and experiment: the late-time mixing growth rate, $α$, measured in most laboratory experiments ($\sim$ 0.06-0.07), is roughly three times the value from idealized direct numerical simulations (DNS, $\sim$ 0.02). The gap's origin remains debated. These properties make RTI a stringent test for a question that matters well beyond RTI: can foundation models trained only on simulations generalise to sparse, messy, and noisy laboratory settings? We finetune Walrus, a foundation model for continuum dynamics, on three or fewer DNS realizations and recover key RTI physics over long rollouts. Applied zero-shot to sliding-barrier laboratory data, the finetuned model leaves the DNS-like regime and enters the observed growth band, having never seen a single experimental sample. These results provide independent, data-driven evidence that initial conditions play a crucial role in the longstanding sim-experiment gap in $α$. The model also generalises zero-shot to stable stratification, a buoyancy regime absent from training, correctly slowing mixing-layer growth. Together, our results show that foundation models can generalise well beyond their training data, predicting laboratory behavior and unseen physical regimes, opening new ways to probe longstanding simulation-experiment gaps.

Electroencephalography (EEG) supports a variety of brain-computer interface (BCI) tasks ranging from brain-state monitoring to human-LLM interactions. EEG foundation models are emerging, but evaluation remains fragmented due to heterogeneous datasets and nconsistent task protocols. Here, we introduce OmniEEG-Bench, a unified benchmark and downstream task roadmap for EEG foundation models (FMs). It organizes evaluation of EEG FMs into six task families spanning (i) signal reliability, (ii) biometrics and disease, (iii) consciousness and state, (iv) cognition and emotion, (v) naturalistic stimulus decoding, and (vi) motor and interaction, introducing a new generation of tasks not systematically benchmarked in prior EEG FM work. OmniEEG-Bench standardizes model deployment, task definitions, and metrics through a task-card specification, and unifies 54 EEG datasets with consistent evaluation protocols. We benchmark 10 representative EEG foundation models and report a leaderboard that covers diverse evaluation settings. Both pretraining dataset diversity and model size are significantly associated with better average ranks across datasets, revealing scaling-law behavior in EEG foundation models (Figure 1). These results suggest that scaling EEG foundation models requires not only larger architectures but also broader and more diverse pretraining data. The benchmark code is available at https://github.com/ncclab-sustech/omni-eegbench.git.

Brain foundation models (BFMs) are self-supervised Transformers pretrained on fMRI data. We posit that these models should capture each subject's cognitive performance from their fMRI signal. Yet across three state-of-the-art BFMs and every readout we test, they predict cognition worse than a linear regression from the $\sim$80K parameters of the functional connectivity matrix (FC). The gap widens with scale: BrainLM's 650M model predicts cognition worse than its 111M. We attribute this to a \textbf{variance allocation problem}: BFM pretraining captures the variance components that dominate fMRI but not the higher-order structure that predicts cognition. Our per-cumulant analysis of the reconstructed signal shows that the second-order covariance is partially preserved, while the third-order co-skewness tensor is largely destroyed. To recover what BFMs lose, we design a linear pipeline that projects the fMRI signal into the subspace that best preserves its co-skewness and computes FC there. This \textbf{exceeds raw FC and every pretrained BFM} on every dataset and parcellation we test, outperforming prior state-of-the-art under controlled evaluation \textbf{with no pretraining and no GPU}. We \textbf{recover the raw-FC ceiling on BrainLM's forward pass} by finetuning with a loss targeted at this same subspace. This shows that the bottleneck is the pretraining objective, not the architecture or the model size.