NeurIPS 2026 论文列表

Large Vision-Language Models (LVLMs) have shown remarkable potential for embodied AI, yet a critical bottleneck persists: the lack of fine-grained spatio-temporal alignment between high-level textual reasoning and low-level visual grounding. Existing benchmarks fail to recognize the critical importance of such fine-grained alignment for embodied tasks. To address this gap, we introduce ECLBench, a comprehensive benchmark designed to systematically assess spatio-temporal embodied understanding across two primary capabilities: Embodied Cognition (Object Cognition and Spatial Cognition) and Embodied Localization (Grounding and Pointing). Together, these four pillars encompass 21 specialized sub-capabilities, supported by 3,616 egocentric video clips (577,998 frames) and 12,000 meticulously curated open-ended questions. Extensive evaluations on ECLBench reveal that general LVLMs excel in semantic understanding but struggle with precise spatio-temporal grounding, while existing embodied agents often sacrifice generalization for specialization. Guided by the benchmark's findings, we develop ECLBrain, a baseline model that discretizes continuous spatial coordinates into integer tokens. This simple yet effective design bridges textual semantics and spatio-temporal vision, demonstrating a promising path toward unified embodied reasoning. ECLBench and ECLBrain together establish a rigorous, multi-dimensional standard for diagnosing current limitations and guiding future research toward spatially-aware and temporally-consistent embodied intelligence.

Current video foundation models, including the strongest self-supervised models such as V-JEPA2, fail to capture how humans organize social information in dynamic scenes. For example, across a range of diverse vision models tested, none were able to predict human similarity judgments to social video clips as well as a sentence embedding model of the caption text (MPNet). We show this gap in vision model performance can be closed by a compact behavioral supervisory signal. We introduce behavioral geometric supervision (BGS): a hybrid objective that constrains local and global pairwise embedding geometry to match the relational similarity structure across videos. We apply this method using a new human similarity dataset, containing 49,484 odd-one-out judgments from 250 naturalistic social video clips, and low-rank adaptation across four ViT backbones (V-JEPA 2/2.1, TimeSformer, VideoMAE, and CLIP). We find that one of the best fine-tuned models, V-JEPA 2.1, nearly triples in performance compared to the pre-trained baseline and reaches close to the noise ceiling, exceeding the strongest sentence-embedding baseline. In addition, finetuned models (i) capture unique variance in human judgments that caption-based language embeddings do not, (ii) develop interpretable social-affective attributes (valence, arousal, and dominance) despite never being trained on any of these attributes, (iii) zero-shot transfer to a separate dataset of out-of-distribution abstract social interactions, and (iv) shift spatial attention from scene context to socially informative regions (faces, gaze, and interacting bodies). A matched language-distillation control fails to reproduce these gains, ruling out caption transfer as the mechanism. Our results show how a modest amount of human behavioral data can steer video models toward human-like social visual understanding.

Behavioral Foundation Models (BFMs) are an emerging paradigm in reinforcement learning, playing a role analogous to large language models in natural language processing: they have shown remarkable versatility, enabling zero-shot performance, fast imitation, and online adaptation, all by exploiting the structure of a latent space. In this work, we investigate whether the latent behavioral space induced by BFMs can serve as an effective search space to discover large repertoires of behaviorally diverse and high-performing policies through Quality-Diversity (QD) methods. While QD methods generally search directly in high-dimensional policy parameter space, in this paper, we present BFM-QD, a framework that performs QD search in the compact latent space of a BFM. We further show that the BFM-QD framework provides a closed-form, gradient-free policy improvement operator that approximates a policy gradient update, but requires no critic training and no backpropagation. Across continuous-control benchmarks spanning dense locomotion, sparse navigation, and contact-rich manipulation, BFM-QD consistently outperforms parameter-space baselines, with particularly stark gains in sparse and deceptive settings, where all existing QD methods collapse to near-zero performance. These results show the effectiveness of the BFM-QD framework, benefiting from the synergy between dimensionality reduction of the search space and offline pretraining from diverse behavioral data. This positions BFMs as a general-purpose backbone for QD optimization, extending their utility beyond zero-shot task solving to the discovery of diverse behavioral repertoires.

Emotions are expressed on a continuum and are naturally described in the valence-arousal (VA) space. Existing VA-based image editors are typically training-based and domain-specific, limiting them to VA regions seen during training and making them brittle across datasets and domains. Meanwhile, despite producing high-quality edits, state-of-the-art foundation image editors are not continuous affective controllers: non-linear emotion representations in text encoders makes prompt interpolation poorly calibrated, so accurate and continuous VA control remains unreliable. To fill this gap, we introduce Barycentric Guidance, a training-free framework that turns pretrained foundation image editors into continuous affective controllers over VA, requiring no finetuning, extra data, or model changes. Our method maps any target VA point to simplex-based guidance weights, enabling VA controllability in three aspects: (i) accurate targeting of VA states, (ii) broad coverage across the VA plane, and (iii) continuous control. Across three face datasets, we improve VA target accuracy by 25–29\%, expand VA-plane coverage by 28–50\%, increase expression diversity by 29–49\%, over the strongest baseline. Unlike prior domain-specific methods, our training-free approach exploits the foundation model’s inherent versatility, enabling generalization beyond human faces to animals, artworks, and complex scenes within a single framework. Code will be made publicly available.

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), which drops discretization and pretrains 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 ($\approx 307$\,h), a fraction ($\approx 3.3$\%) of that required by existing EEG foundation models. It also produces synthetic EEGs that two board-certified neurologists cannot distinguish from real EEGs (Cohen's $\kappa = -0.096$).

Parameter-efficient fine-tuning of quantized large language models is dominated by QLoRA, which combines 4-bit weight quantization with FP16 low-rank adapters. We show the FP16 adapter precision is unnecessary: a single binary bit per adapter parameter, combined with a learned per-layer scale $s$ and a Scaled Straight-Through Estimator, matches or exceeds QLoRA across a broad range of benchmarks while reducing adapter storage by $14$--$16\times$. We introduce two variants--NormScale, which clips STE gradients to the unit ball, and LearnThresh, which makes the binarization threshold learnable---unified under a framework we call BQ-LoRA. Our main theoretical contribution is a tight characterization of the BQ-LoRA hypothesis class: it has bounded entrywise norm, a high-probability spectral norm bound that is $\sqrt{r}$ times sharper than the worst case, a generically full-rank update, and consequently a Rademacher complexity that yields an explicit $\tilde{O}(s_{\max}\sqrt{d_{\text{in}} d_{\text{out}}/n})$ generalization bound. None of these properties hold a priori for FP16 LoRA. On Qwen2.5 (1.5B/7B/72B) and LLaMA-2 (7B/70B), BQ-LoRA matches or beats QLoRA and LoftQ at $\sim 1/16$ the adapter footprint; the predicted regularization signature appears empirically in dataset ablations and training-loss curves. We extend Punica's SGMV kernel with a binary inner-product primitive (BQ-SGMV) and prove a cache phase-transition result: shrinking adapters $16\times$ enlarges the HBM cache from $2{,}500$ to $39{,}000$ adapters, eliminating misses under skewed workloads.

Tabular Foundation Models (TabFMs) produce context-aware embeddings that are increasingly stored, shared, and reused for retrieval, clustering, and cross-institutional data exchange. These embeddings are often treated as privacy-preserving substitutes for raw tabular records, yet their attribute-level information content remains poorly understood. We present a systematic privacy audit of TabFM embeddings: given an embedding and varying amounts of side information, how accurately can sensitive attributes be recovered? We introduce *Cascade Probing*, a sequential probing method that measures recoverability while accounting for inter-attribute dependencies, significantly outperforming joint probing and optimization-based inversion baselines. Across four TabFMs (TabPFN, TabDPT, Mitra, TabICL) and eight datasets, we find that even without any side information, 65-95\% of sensitive attributes can be recovered from embeddings alone. More critically, recoverability exhibits a *Privacy Cliff*: revealing a single known attribute sharply increases the recoverability of all remaining attributes. This phenomenon generalizes across model architectures and diverse domains, and is not eliminated by standard embedding transformations including noise injection, PCA, and knowledge distillation. These findings suggest that we need more formal privacy techniques for protecting sensitive information inside TabFM embeddings.

Most black-box optimization methods require extensive hyperparameter tuning, often limiting their ability to generalize across different optimization domains. Foundation models for black-box optimization that learn optimization principles from a large collection of optimization trajectories offer a promising alternative, with the potential to outperform manually designed methods across diverse problem classes. However, prior work has either relied on non-public datasets or on purely synthetic data, limiting reproducibility and generalization to real-world problems. As a result, progress in this area has been constrained by the lack of large-scale, real-world, publicly available pre-training data. We introduce BBO-Pile, the first open-source dataset comprising over 500K optimization trajectories evaluated across 3095 different black-boxes for different optimizers which represents by far the largest public dataset for this task. Using this dataset, we train a family of foundation models at multiple scales, ranging from 2M to 80M parameters and from 200M to 2B training tokens, and study their scaling behavior with respect to compute. Our results demonstrate that large-scale pre-training is a viable and effective approach to imitate black-box optimization methods, paving the way for future research in this direction.

Modern foundation models are trained through multistage pipelines, typically beginning with unsupervised pretraining, followed by supervised fine-tuning (SFT) on step-by-step solutions, and reinforcement learning (RL) with bulk feedback. Empirically, the design of these stages, particularly the compute allocated to each, strongly affects performance. Yet our theoretical understanding of multistage training remains limited. Here, we introduce a simple, analytically tractable model of multistage training. The model reproduces several qualitative phenomena: SFT and RL can be complementary, pretraining quality imposes a performance ceiling with power-law scaling, optimal compute allocation varies with data quality and pretraining level, and RL can selectively amplify task-relevant pretrained skills. Our model distills the complex behaviors into key ingredients, allowing clearer understanding of how and where they arise. Overall, our model provides a theoretical starting point for explaining the phenomenology of multistage training pipelines in the compute-limited regime.

Ambient intelligence seeks to continuously understand human presence, activity, and physiology in physical spaces, enabling smart environments, health monitoring, and human-computer interaction. WiFi infrastructure offers a ubiquitous, always-on, and privacy-preserving sensing substrate across billions of IoT devices. Yet ambient sensing with WiFi remains largely fragmented, with most systems relying on task-specific models, labeled data collection, and customized training pipelines. We present AmbientFM, the first foundation model for ambient sensing through ubiquitous WiFi signals. AmbientFM is pre-trained on 9.2 million unlabeled Channel State Information (CSI) samples collected over 439 days from 20 commercial device types deployed in real-world environments. It learns transferable wireless representations through contrastive learning, masked reconstruction, and physics-informed objectives tailored to wireless signals. With lightweight adaptation, a single backbone transfers to 9 downstream tasks, achieving above 0.90 AUROC on all classification benchmarks and supporting dense spatial reconstruction. AmbientFM reduces the need for labeled data and task-specific design, enabling scalable ambient intelligence on existing wireless infrastructure.

Although Large Language Models (LLMs) achieve strong alignment through supervised fine-tuning and reinforcement learning from human feedback, the alignment is often fragile under subsequent fine-tuning. Existing explanations either attribute alignment fragility to gradient geometry or characterize it as a distributional shift in model outputs, yet few provide a unified account that bridges parameter-space learning dynamics with function-space alignment behavior during fine-tuning. In this work, we introduce a tractable alignment score and derive its closed-form update during fine-tuning, yielding a unified framework for alignment dynamics. Our analysis decomposes alignment updates into two competing components: a Rebound Force, governed jointly by the current alignment state and the narrowness of model distribution, and a Driving Force, determined by how the training distribution aligns with outcome-conditioned posteriors over aligned and non-aligned completions. This decomposition explains why prior alignment can be reversed by later fine-tuning and why narrower posterior structure strengthens such reversal. Moreover, our framework predicts a Rehearsal Priming Effect: prior alignment leaves a latent posterior imprint that amplifies the effective Driving Force upon re-exposure, leading to faster re-alignment. We validate these predictions across safety alignment, emergent misalignment, and sentiment settings, demonstrating consistent alignment reversal and accelerated re-alignment under re-exposure. In addition, controlled experiments in safety alignment confirm the predicted dependence of rebound strength on posterior narrowness. Together, these results provide a unified dynamical perspective on how alignment is disrupted and reactivated during LLM fine-tuning.

Engineering LLMs to accelerate life sciences research requires a robust alignment with biomedical knowledge. We observe that biomedical text exhibits a fundamentally different uncertainty structure from general text: dense low-confidence runs encode epistemic knowledge gaps (dense causal chains, rare entities) rather than the sparse aleatoric stylistic variation typical of general text. Based on this discovery, we propose Balanced Fine-Tuning (BFT), a dual-scale post-training method that combines group-normalized token reweighting with sequence-level reallocation toward knowledge-dense samples exhibiting dense epistemic uncertainty. Across medical evaluation, biological reasoning, sparse-reward RL, and biological representation tasks, BFT provides more consistent gains than SFT and DFT under a shared training setup. When replacing the default closed-source backbones in GeneAgent (GPT-4o) and VCWorld (Gemini-2.5-Flash), the BFT-aligned 70B model delivers stronger performance across biological process reasoning and chemical perturbation prediction. Critically, all BFT variants further improve after subsequent GRPO with sparse rewards, while SFT and DFT degrade, suggesting that epistemic-aware post-training provides a more robust policy initialization. Beyond text generation, BFT-aligned LLMs produce more accurate and professional biomedical profile texts; after encoding these profiles with a text embedding model, the resulting representations support gene-level, cell-level, and perturbation-response tasks, suggesting that BFT-enhanced generation can facilitate biological representation and, in turn, broader biomedical downstream tasks.

Federated Low-Rank Adaptation (FedLoRA) enables efficient fine-tuning of large language models by exchanging low-rank adaptation parameters across clients. Several studies mainly focus on aggregation bias and design improved aggregation schemes. We argue that aggregation bias is not the fundamental bottleneck of FedLoRA. Even under ideal aggregation, heterogeneous data induce directional conflicts among client-specific LoRA updates. We show that these conflicts arise from misaligned low-rank adaptation subspaces, leading to destructive interference during collaboration. To diagnose this phenomenon, we introduce a geometry-aware metric that characterizes client updates at the subspace level. Our analysis reveals that subspace conflicts intensify with increasing heterogeneity and are concentrated in the column spaces of LoRA updates. Based on this insight, we propose ALIGN, a geometry-aware federated adaptation framework that aligns client LoRA subspaces during aggregation and constrains local updates to mitigate conflicts. Comprehensive empirical results confirm that mitigating subspace misalignment is crucial for FedLoRA, and that ALIGN consistently improves downstream performance over existing methods. The code is available for anonymous access at https://anonymous.4open.science/r/ALIGN-7C41.

Low-Rank Adaptation (LoRA) reparameterizes a weight update as a product of two low-rank factors, but the Jacobian $J\_\mathcal{G}$ of the generator mapping the factors to the weight matrix is rank-deficient, so the factor-space preconditioner $J\_\mathcal{G}^\* \mathcal{F}\_t J\_\mathcal{G}$ induced by any ${W}$-space preconditioner $\mathcal{F}\_t$ is singular, and consequently the standard chain rule cannot be uniquely inverted to map a preconditioned ${W}$-space direction back to a factor-space update. We cast existing LoRA optimizers in a unified framework parameterized by two choices: (i) which invertible surrogate for $J\_\mathcal{G}^\* \mathcal{F}\_t J\_\mathcal{G}$ to use, and (ii) which $\mathcal{F}\_t$ on ${W}$ to use. Existing methods occupy four families along these axes: factor-space adaptive updates, block-diagonal surrogates for $J\_\mathcal{G}^\* J\_\mathcal{G}$, Frobenius-residual pseudoinverse methods, and Riemannian manifold constraint. Within this design space, a gradient-statistics-aware $\mathcal{F}\_t$ paired with a closed-form factor-space solve at $\mathcal{O}((m+n)r)$ memory remains underexplored. We propose AdaPreLoRA, which fills this gap by adopting the Adafactor diagonal Kronecker preconditioner $\mathcal{H}\_t$ on ${W}$ and selecting from the resulting factor-space solution family the element minimizing an $\mathcal{H}\_t$-weighted imbalance between the two factor contributions; by construction, the resulting factor update is the closest LoRA approximation to the preconditioned ${W}$-space direction under the $\mathcal{H}\_t$-weighted norm. Across GPT-2 (E2E), Mistral-7B and Qwen2-7B (GLUE, ARC, GSM8K), and diffusion-model personalization, AdaPreLoRA is competitive with or improves over a representative set of LoRA optimizers while keeping peak GPU memory at the LoRA optimizer level.

Zeroth-order (ZO) optimization enables large-language-model fine-tuning without storing backpropagation activations, while LoRA supplies compact trainable adapters. Combining them creates a rank paradox: increasing LoRA rank improves adapter capacity, but standard two-point ZO either perturbs a rank-dependent number of coordinates or, under atomwise updates, can make the finite-difference signal unobservable. This paper shows that the bottleneck is a measurement-topology problem rather than a need for an external subspace. LoRA already decomposes into matched rank-$1$ atoms, each a complete factor-coordinate block of dimension $d_\text{out}+d_\text{in}$. Querying one atom per step keeps the stored adapter rank $r$ while removing $r$ from the single-query perturbation dimension. The naive atomwise query is still miscalibrated: if it inherits canonical LoRA scaling $\alpha/r$, the active finite-difference signal shrinks as $1/r$ and the active finite-difference signal-to-noise ratio (FD-SNR) as $1/r^2$, producing directional collapse under a fixed residual evaluation-noise floor. AR1-ZO pairs alternating rank-$1$ atom queries with topology-aware scaling $\gamma=\alpha r$, restoring rank-invariant active signal without auxiliary bases, activation hooks, curvature estimates, or extra forward queries. Theory proves atom minimality, rank-independent active query dimension, directional collapse and restoration, and the remaining rank dependence as an amortized coverage cost. Experiments on OPT and Qwen3 models validate the signal mechanism and show that AR1-ZO makes high-rank LoRA effective among matched-budget ZO methods under the standard two-forward-pass query budget.

Foundation models are increasingly deployed as agentic systems that perceive context and act autonomously, raising the need to evaluate their moral behavior as active decision-makers. Existing benchmarks, however, largely ask models to judge human actions from third-person textual vignettes, leaving untested how models behave when they themselves become the acting agent. We introduce ABHBench, a multimodal benchmark built from \emph{Detroit: Become Human}, an interactive narrative game, to evaluate how foundation models make moral decisions in first-person human-android dilemmas. ABHBench places foundation models in branching narrative scenarios with visual observations, dialogue, memory, and action choices. Inspired by Asimov's Three Laws of Robotics, our framework measures Harmlessness, Obedience, and Self-preservation to capture the core tensions in human-AI relations, along with Autonomy, Stability, and Certainty to assess decision-making patterns. Across 12 foundation models, we find that moral behavior changes substantially under situated agentic context: 75\% of models shift significantly between a text-only abstraction and the matched multimodal DBH scenario, and reasoning-enhanced models show no monotonic gains in consistency or human-oriented behavior. These results highlight the necessity of evaluating moral behavior in situated agentic contexts rather than relying on third-person judgment alone.

Graph domain adaptation (GDA) has emerged as an important problem in graph machine learning when the distribution of the source graph used for training differs from that of the target graph used for testing. While much of the prior work on GDA has focused on aligning node representations across source and target domains, recent studies show that such approaches can be suboptimal in the presence of graph structure shift, where the underlying connection patterns between nodes change across domains. In this work, we develop a unified pairwise distribution matching framework for mitigating conditional structure shift (CSS), a specific form of graph structure shift in which conditional edge distributions vary across domains. The framework recovers existing GDA methods as instances of moment matching and motivates PLSA, a new likelihood-based method that uses calibrated probabilistic predictors on connected target node pairs. Theoretically, we establish finite-sample guarantees for both the likelihood-based and moment-matching estimators under the contextual stochastic block model. Our analysis uses Bernstein-type concentration bounds for edge-weighted U-statistics, leading to error bounds that reflect both the effective number of observed edges and the conditioning of the corresponding likelihood or moment-matching problem. We complement our theoretical results with empirical studies that demonstrate the effectiveness of the proposed framework.

Foundation-model-based predictors of perturbation response are evaluated primarily by aggregate accuracy, yet lack principled per-prediction reliability criteria. We audit the state-transition (ST) decoder~\citep{adduri2025predicting} shared by recent predictors across three encoders and four cell lines under CRISPRi perturbation, and identify a rank-$1$ attention bottleneck whose leading direction maps in gene space to the same cycling/p53 stress-response program in every cell line tested. The bottleneck implies a distance-defined reliability boundary: per-prediction quality degrades smoothly with the gene-space distance from a model's predicted perturbation response ($\Delta$, the predicted change relative to control) to its nearest training $\Delta$, and biological-family identity, not training-pool size, drives held-out quality. We propose \emph{predicted-NN-distance}, this same distance computed from model output alone, as a per-prediction trust signal that tracks per-perturbation quality at Spearman $|\rho| \approx 0.81$ (oracle $0.88$) and admits a single pooled threshold that transfers across cells and encoders under leave-one-out evaluation. The signal outperforms a simpler predicted-magnitude baseline in most cases. Per-prediction reliability is bounded by a low-rank decoder geometry and by gene-space coverage of training perturbations, not by aggregate benchmark accuracy.

Graph clustering is a fundamental task for understanding graph-structured data without labels, yet classical end-to-end methods require retraining on each dataset, limiting their generalization ability. Although Graph Foundation Models (GFMs) enable transferable learning across diverse graph tasks, they are not directly applicable to graph clustering. This limitation stems from two key factors: the inability to learn mutually beneficial knowledge on both the node and graph levels, and the lack of a completely unsupervised clustering-oriented design. To bridge this gap, we propose a **G**raph **F**oundation **M**odel for **U**nified **C**lustering (GFM-UC), which pre-trains on the source domain and only requires fine-tuning for fast clustering in the target domain. The entire learning process does not require any labels. Specifically, in the pre-training stage, we select learnable high-confidence unified subgraph prototypes from the source domain. And then we generalize them to train the clustering network collaboratively. In the fine-tuning stage, we align the target domain with the source domain in a distribution-aware manner to further refine the clustering head and obtain more well-separated cluster boundaries. Extensive experiments on five datasets demonstrate that our method significantly outperforms existing state-of-the-art approaches.