NeurIPS 2026 论文列表

Diagnosis of psychiatric disorders from functional magnetic resonance imaging (fMRI) is hindered by low-resource phenotypes, high inter-subject variability, and weakly interpretable multimodal fusion. We propose a prior text-informed gated attention framework (PTGA) that converts region-of-interest (ROI) level neuroimaging measures into statistically grounded textual priors via percentile-based soft binning. PTGA combines hemodynamic response function (HRF) aware spatiotemporal encoding with static graph topology modeling, and unifies variable-length fMRI time series through masked global temporal aggregation. It further uses text semantics as an inductive bias for prior-guided gated fusion, improving robustness under cross-modal noise. We compare PTGA with eight strong graph and multimodal baselines under a unified protocol. PTGA achieves up to 10.46\% and 8.01\% absolute improvements in autism spectrum disorder (ASD) and major depressive disorder (MDD) diagnosis, respectively, outperforming SOTA baselines. The learned gating also provides biologically plausible evidence: MDD related signals are concentrated in sensorimotor/visual/default-mode systems, while ASD related evidence is centered on the cerebellar network with joint contributions from default-mode, limbic, and dorsal-attention networks.

Evaluating physical reasoning in video models is difficult because absolute motion measurements depend on frame rate, object scale, and camera calibration, all of which are often ambiguous or unavailable in generated video. We propose a different approach. When two objects in the same scene obey the same physical law, their motions must satisfy predictable relationships, and these relationships hold independent of calibration. We introduce Principia, a benchmark that evaluates Newtonian physics through relational consistency between paired objects. Principia spans eight phenomena—gravity, restitution, friction, rotational inertia, projectile motion, momentum, pendulum, and mass-spring oscillation—across translational, rotational, collisional, and oscillatory dynamics, using real-world scenes recorded under controlled protocols. We also introduce a calibration-independent consistency score that quantifies physical violation directly in image space. Across thousands of generations from five state-of-the-art video generators, no model exceeds 0.42 on Principia despite all scoring above 0.7 on VBench. Vision-language models show inverse failure profiles, succeeding where generators fail and vice versa.

Personalizing large language models is essential for user-centric applications, yet remains challenging under sparse user histories, privacy constraints, and limited on-device resources. Existing approaches either rely on prompting or retrieval without adapting model parameters, train and store separate adapters for each user, or use coarse group-level adapters that cannot fully capture individual variation. We propose $\textbf{PriSM}$ ($\textbf{Pri}$or-guided $\textbf{S}$hared-basis $\textbf{M}$ixture Personalization), a lightweight framework that models personalization as posterior-inspired update over mixtures of shared LoRA bases. PriSM first learns reusable basis adapters that capture group-level adaptation patterns, then uses a cluster-conditional Dirichlet prior and a user-conditioned hypernetwork to infer posterior mixture weights from sparse user profiles. The resulting posterior mean synthesizes a personalized LoRA update in a single forward pass, allowing the model to rely on group-level priors when user evidence is limited and to adapt toward user-specific behavior when sufficient evidence is available. This design enables scalable, privacy-preserving, and on-the-fly personalization without per-user training or per-user adapter storage. Code is available at \url{https://anonymous.4open.science/r/PriSM-D705}.

Prefix-tuning is a parameter-efficient fine-tuning method for large language models (LLMs) that steers model outputs by prepending task-specific prefix representations to the input sequence. Despite its practical success, the theoretical understanding of the expressiveness of prefixes remains incomplete. In particular, the following question remains open: {\it Given a pretrained autoregressive Transformer such as the GPT series, for any input-output sequence pair $(\mathtt{x},\mathtt{y})$, does there always exist a prefix $\mathcal{S}$ such that prepending $\mathcal{S}$ to $\mathtt{x}$ induces the model to generate the target output $\mathtt{y}$?} In this work, we prove that under the assumption that $\mathtt{y}$ is in the {\it range} of the pretrained Transformer, such a prefix $\mathcal{S}$ exists and can be explicitly computed by solving multiple linear systems. We further show that this assumption is necessary, in the sense that such a prefix may not exist when the assumption fails, thereby providing a complete answer to the above question. Our analysis is constructive and the obtained theoretical results hold under fairly general assumptions on model architecture. As a notable example, the whole GPT-2 architecture satisfies our model assumptions.

Weight-space post-training quantization (PTQ) must choose finite formats, granularities, quantizer families, transformations, and bits before the completed quantized model reveals its output-distribution drift. Existing PTQ methods predict important pieces of this degradation, including reconstruction error, Hessian sensitivity, transformation effects, and downstream loss, but these pieces are usually scored after fixing the quantization geometry or inside separate configuration families. We formulate weight-space PTQ as pre-deployment configuration selection using priced layer-output error. Each admissible layer configuration is treated as an error generator with a deployment cost, which induces a layer-output error covariance $\boldsymbol{\Sigma}_l(\alpha_l)$, and the full-precision model prices that covariance by downstream curvature, $\hat{\rho}_l(\alpha_l)=\frac{1}{2}\operatorname{Tr}\left(\hat{\mathbf{H}}_l\,\hat{\boldsymbol{\Sigma}}_l(\alpha_l)\right)$. The price follows from full-precision-to-quantized forward KL, whose first-order term cancels at the reference model. It turns reconstruction and diagonal scores into reduced proxies that drop price factors, while finite formats, codebooks, granularities, and equivalent transformations become comparable candidates through the covariances they induce and the costs they pay. A trace reduction then yields a calibration-time price table and a budgeted price-guided selector, making fixed-geometry bit allocation a special case rather than the organizing problem.

We study posterior contraction rates for sparse Bayesian Kolmogorov-Arnold networks (KANs) over anisotropic Besov spaces, providing a statistical foundation of KANs from a Bayesian point of view. We show that sparse Bayesian KANs equipped with spike-and-slab-type sparsity priors attain the near-minimax posterior contraction. In particular, the contraction rate depends on the intrinsic anisotropic smoothness of the underlying function. Moreover, by placing a hyperprior on a single model-size parameter, the resulting posterior adapts to unknown anisotropic smoothness and still achieves the corresponding near-minimax rate. A distinctive feature of our results, compared with those for standard sparse MLP-based models, is that the KAN depth can be kept fixed: owing to the flexibility of learnable spline edge functions, the required approximation complexity is controlled through the network width, spline-grid range and size, and parameter sparsity. Our analysis develops theoretical tools tailored to sparse spline-edge architectures, including approximation and complexity bounds for Bayesian KANs. We then extend to compositional Besov spaces and show that the contraction rates depend on layerwise smoothness and the effective dimension of the underlying compositional structure, thereby effectively avoiding the curse of dimensionality. Together, the developed tools and findings advance the theoretical understanding of Bayesian neural networks and provide rigorous statistical foundations for KANs.

As synthetic images become increasingly realistic, reliable synthetic image detection techniques are of pressing need to prevent their misuse. Despite satisfactory in-distribution performance, deep neural network-based synthetic image detectors (SIDs) lack reliability in deployment and often fail in the presence of common covariate shifts, resulting in poor detection accuracy. To avoid the risk caused by potential errors, we adopt a selective classification (SC) strategy by allowing SIDs to abstain from making low confidence predictions. For practicality, we focus on post-hoc methods which perform confidence estimation on a given SID without retraining. However, we show that conventional logit-based confidence score functions (CSFs) exhibit pathological behavior under covariate shifts, leading to SC performance close to or even worse than random guessing. To address this, we propose a simple yet effective SC framework for **Re**liable **S**ynthetic **I**mage **De**tection (ReSIDe). First, we generalize the notion of logits to an SID's intermediate layers from a centroid matching perspective, extending the use of logit-based CSFs to any layer of an SID. Then, we introduce a preference optimization algorithm that aggregates confidence scores extracted from different layers to a final confidence estimate by minimizing an upper bound of the area under the risk-coverage curve (AURC). Extensive experimental results show that ReSIDe significantly boosts the SC performance of various logit-based CSFs under common covariate shifts, achieving up to 69.55% AURC reduction.

Vision Transformers (ViTs) demonstrate strong performance, but their substantial memory footprint and computational overhead necessitate efficient compression techniques such as post-training quantization (PTQ). However, pushing ViTs to low-bit precision incurs significant accuracy degradation. Recent PTQ methods adopt block reconstruction-based optimization with curvature-aware objectives, typically implemented using Fisher-based approximations. These approaches rely on explicit curvature modeling based on diagonal or structured approximations, which still fail to capture cross-dimensional interactions. Moreover, they often involve matrix inversion, leading to numerical instability under ill-conditioned settings. To address these limitations, we propose Gradient-Projected Fisher Approximation for Quantization (GPFA-Q), a block reconstruction-based PTQ framework that avoids explicit curvature matrix construction while capturing off-diagonal interactions. First, we introduce Gradient-Projected Reconstruction (GPR), a reconstruction objective that captures cross-dimensional interactions without explicitly constructing curvature matrices. To further support GPR, we integrate Soft Grid Rounding (SGR), which reduces the mismatch between continuous reconstruction and discrete inference. Extensive experiments demonstrate that our GPFA-Q achieves the state-of-the-art performance in low-bit quantization across diverse vision tasks.

Vision-Language Models (VLMs) for radiology report generation are typically trained on retrospective clinical reports, which suffer from omission noise: clinically present findings are left unreported due to the omission of subtle findings. For example, prior studies show that cardiomegaly may be omitted from ICU chest X-ray reports when the imaging request is focused on monitoring support device placement. As a result, models trained with standard approaches inherit these omissions, learning to under-report findings themselves. We propose PU-DPO, a preference optimization framework that constructs contrastive pairs via editing model responses, producing variants that explicitly mention or omit a specific finding. To prevent omission noise from corrupting the preference signal, we reformulate the objective under a positive-unlabeled (PU) learning framework, treating absent mentions as unlabeled rather than truly negative. Across semi-synthetic experiments and preliminary analyses on real-world chest radiograph benchmarks where adjudicated labels are available, PU-DPO yields consistent gains in detection rates across multiple pathologies without compromising specificity or overall report quality, and is more robust to omission noise than prior approaches.

Positional encoding in transformers is commonly implemented through positional embeddings, attention masks, or bias terms, but formal connections between these mechanisms remain limited. We study attention with positional bias through the lens of locality-sensitive hashing (LSH), focusing on Attention with Linear Biases (ALiBi). We show that the ALiBi bias matrix is the expectation of contiguous block-diagonal binary masks induced by a ``positional LSH'' scheme. The empirical mean of masks sampled from this scheme yields spectral norm and max-norm approximation guarantees with bounded block sizes with high probability. This structural theorem implies a uniform approximation theorem for ALiBi-biased attention: with high probability over the sampled masks, the approximate attention output is accurate simultaneously for all query-key-value inputs and can be computed in near-linear time in the context length, reducing long-context ALiBi to a collection of randomized short-context regular (positionally unbiased) attention operations. Conceptually, this connects positional bias, masks, and positional embeddings in a single formal framework and suggests an approach to efficient ALiBi-biased attention. Experiments on large language models validate our theoretical findings.

Vision-language-action (VLA) policies have shown strong potential for general-purpose manipulation, yet they often fail on novel, out-of-distribution objects whose appearance or geometry deviates from the training distribution. The standard remedy is to collect multi-view teleoperation data for every failure case, but this scales poorly in both cost and time. We introduce Pose6DAug, a failure-driven data augmentation framework that turns a policy’s own successful episodes into targeted demonstrations for its failure modes, without any new data collection. Our key insight is that each successful episode already encodes a physically valid action trajectory together with calibrated multi-view observations. By swapping only the manipulated object while preserving this trajectory, we obtain new and physically grounded demonstrations. However, the challenge lies in synthesizing object swaps, e.g., naive 2D video editing breaks multi-view consistency and physical plausibility, particularly under heavy occlusion and egocentric viewpoints. Our method instead operates directly in 3D, anchoring the target object with an explicit mesh driven by a temporally coherent 6D pose trajectory. This ensures geometrically consistent renderings across all camera views. Fine-tuning a VLA on data augmented by our method improves success rates by 16.5% relative to the state-of-the-art baseline on novel objects, while preserving in-distribution performance. These results show that multi-view and physically consistent augmentation is a practical path to scalable VLA generalization.

Multimodal large language models have achieved remarkable progress, yet their language-dominant designs constrain visual perception and reasoning. Vision-centric multimodal large language models aim to restore visual understanding as the structural foundation, but existing systems rely on static visual expert fusion, leading to redundant computation, weak adaptivity, and limited multimodal transferability. Prior VC-MLLMs scale visual capacity by adding experts, whereas $\mathtt{PolyVision}$ scales visual capacity by allocating experts. We introduce $\mathtt{PolyVision}$, a vision-centric framework for conditional visual scaling in which dynamic visual expert routing assigns image patches to heterogeneous visual encoders. An \textbf{Attentive Router} selectively activates relevant experts for each patch, while a compact \textbf{VisPack} module executes routed representations efficiently and preserves example independence. Together with additional visual pre-alignment, this design turns heterogeneous visual scaling into a practical conditional-computation process. Built on the \texttt{Qwen3} backbone, $\mathtt{PolyVision}$ achieves superior efficiency and scalability on multimodal benchmarks. Experiments show that adaptive expert routing strengthens visual specialization, clarifies practical properties of visual scaling, and improves visual understanding at a favorable trade-off between efficiency and performance.

The inference efficiency of diffusion large language models (dLLMs) is constrained by two challenges: bidirectional attention precludes efficient KV-cache reuse, while increasing decoding parallelism with static confidence thresholds can compromise generation quality. We observe that both challenges arise from a shared phenomenon: as tokens are decoded, their contextual integration through bidirectional attention causes token representations to drift (evolve) across decoding steps. This insight motivates Polestar, a training-free inference framework that uses token representation drift as a unified signal to jointly address both challenges. Polestar comprises two components: Polestar-Cache, which identifies stale KV-cache positions via drift and performs sparse KV-cache refreshes to enable efficient reuse, and Polestar-Commit, which detects sharp drift events to reliably identify commit-ready tokens. Across mathematics and coding benchmarks on several dLLM families, Polestar sets a new state of the art on the accuracy-throughput Pareto frontier, achieving up to 10.73% accuracy improvement, up to 3.7x higher throughput, and high decoding parallelism of 3.67 tokens per forward pass over existing baselines.

Polarization imaging provides rich physical cues beyond RGB imaging, yet acquiring such information typically requires specialized polarization-sensitive hardware that increases system cost and deployment complexity. Recent methods have proposed to infer polarization information directly from standard RGB images, providing a hardware-free alternative. While useful for describing relative polarimetric structure, these representations do not fully assess whether estimated images preserve the radiometric scale and intensity-dependent physical information needed for full Stokes reconstruction. To address this limitation, we introduce PolarScale, a physics-grounded benchmark for radiometrically consistent RGB-to-Stokes estimation. Instead of treating polarization inference from RGB images as conventional image-to-image translation, PolarScale evaluates how scale-independent descriptors, scale-dependent Stokes components, and radiometric scale jointly support physically meaningful Stokes reconstruction. Building upon this formulation, we propose three prediction strategies, namely direct, joint, and decoupled prediction, to study how scale-independent and scale-dependent polarimetric cues should be modeled across representative restoration-based and generative frameworks. Our analysis reveals that restoration-based frameworks outperform generative ones for radiometrically consistent polarization estimation (e.g., 28.61 vs. 26.14 dB PSNR). Furthermore, we find that jointly modeling scale-independent and scale-dependent parameters yields more stable physical representations than direct scale-dependent prediction (e.g., 28.61 vs. 26.81 dB PSNR). Overall, PolarScale provides a physics-grounded benchmark foundation for evaluating whether vision models can recover radiometrically consistent polarimetric signals from standard RGB images.

Many projection-based imaging modalities reconstruct unknown objects from collections of noisy projections related by unknown in-plane rotations, making effective estimation dependent on rotationally consistent alignment and aggregation. In cryogenic electron microscopy (cryo-EM), the extremely low signal-to-noise ratio (SNR) makes such multi-image integration essential. We introduce the *polar transformer*, a permutation-invariant architecture for *sets* of images built around an $\mathrm{SO}(2)^K$-equivariant *angular attention* mechanism that jointly aligns and aggregates information across images. This architecture leverages a stable polar representation in which we define polar CNN feature extractors and evaluate attention weights over all relative angular shifts efficiently via FFT-based correlations, guaranteeing equivariance to independent in-plane rotations of each image in the set. We evaluate the polar transformer on simulated cryo-EM images using denoising as a quantitative proxy. Across different regimes, polar transformers outperform strong single-image and image-set baselines, achieving a reduction of up to 40% in relative MSE at an SNR of 0.02 and improving downstream ab initio reconstruction resolution by 20%.

Poisoning attacks can compromise the safety of large language models (LLMs) by injecting malicious documents into their training data. Existing work has studied pretraining poisoning assuming adversaries control a percentage of the training corpus. However, for large models, even small percentages translate to impractically large amounts of data. This work demonstrates for the first time that poisoning attacks instead require a near-constant number of documents regardless of dataset size. We conduct the largest pretraining poisoning experiments to date, pretraining models from 600M to 13B parameters on chinchilla-optimal datasets (6B to 260B tokens). We find that 250 poisoned documents similarly compromise models across all model and dataset sizes, despite the largest models training on more than 20 times more clean data. We also run smaller-scale experiments to ablate factors that could influence attack success, including broader ratios of poisoned to clean data and non-random distributions of poisoned samples. Finally, we demonstrate the same dynamics for poisoning during fine-tuning. Altogether, our results suggest that injecting backdoors through data poisoning may be easier for large models than previously believed as the number of poisons required does not scale up with model size—highlighting the need for more research on defences to mitigate this risk in future models.

Perturbation experiments are central to understanding cellular mechanisms, but remain costly and sparse, motivating prediction of gene expression responses for unobserved conditions. A promising recent direction leverages large language models (LLMs) as “virtual cell” simulators—using stepwise, knowledge-grounded mechanistic reasoning to infer differential expression—pointing toward an interpretable, knowledge-driven paradigm that transcends purely data-driven approaches. However, we find that plausibility is not prediction: despite producing biologically plausible explanations, these methods fail to capture perturbation-specific effects — systematically overestimating differential expression, often underperforming a simple gene‑frequency baseline in aggregate evaluations, and collapsing to chance-level performance at the per-gene level. This reveals a reliance on intrinsic gene response tendencies rather than true perturbation reasoning. We trace this failure to how evidence is presented: existing methods evaluate perturbation–gene pairs in isolation, without exposing how related perturbations differ in their effects on the same gene. To address this limitation, we introduce **CORE** (**C**ontrastive **O**rganization of **R**elational **E**vidence), which reframes prediction as a comparison task by organizing evidence into positive and negative outcomes from related perturbations. Using a biomedical knowledge graph for evidence retrieval, CORE improves calibration and substantially boosts perturbation-specific prediction in both LLM-based and non-LLM settings: for example, on drug-perturbation data, CORE-Reasoning improves Qwen3.5-9B aggregate metrics by up to 28.6\%, while on generic perturbation data, CORE-Voting raises macro-per-gene AUROC from chance to 0.711. This highlights contrastive evidence organization as essential to reliable LLM-based perturbation reasoning.

Recent advances in diffusion-based text-to-image generation have demonstrated promising results through visual condition control. However, existing ControlNet-like methods struggle with compositional visual conditioning - simultaneously preserving semantic fidelity across multiple het erogeneous control signals while maintaining high visual quality, where they employ separate control branches that often introduce redundant guidance during the denoising process, leading to structural distortions and artifacts in generated images. To address this issue, we present PixelPonder, a novel unified control framework, which allows for effective control of multiple visual conditions under a single control structure. Specifically, we design a patch-level adaptive condition selection mechanism that dynamically prioritizes spatially relevant control signals at the subregion level, enabling precise local guidance without global interference. Additionally, a time-aware control injection 017 scheme is deployed to modulate condition influence according to denoising timesteps, progressively transitioning from structural preservation to texture refinement and utilizing the control information from different categories to promote finer image generation. Extensive experiments demonstrate that PixelPonder surpasses previous methods across different benchmark datasets, showing superior improvement in spatial alignment accuracy while maintaining high textual semantic consistency.

Representation alignment (REPA) accelerates diffusion transformer training, but its alignment targets are almost exclusively semantic encoders such as DINOv2 and CLIP. Recent analysis points to spatial structure, not global semantics, as the carrier of the alignment effect, yet dense-prediction foundation models trained to predict that structure remain overlooked as REPA targets. In pixel-space diffusion, SAM2, Depth Anything v2, and Metric3D v2 each outperform the DINOv2-only GenEval baseline, with the two geometric teachers leading the segmentation teacher. A flat sum of all four teachers, however, lands below the best single geometric teacher, semantic teachers reward viewpoint invariance and instance identity while geometric teachers reward metric layout and surface orientation, and forcing both kinds of gradients through one denoiser projection collapses them into a shared subspace. We introduce PixelDense, which routes DINOv2 and SAM2 through a semantic projection stream, routes Depth Anything v2 and Metric3D v2 through a geometric projection stream, and adds a weight-space orthogonality penalty that keeps the two streams in disjoint subspaces. All four teachers are frozen during training and dropped at inference. Applied to PixelGen and DeCo with a single recipe, PixelDense improves GenEval, DPG-Bench, and HPS v2.1, raises PixelGen-XXL's GenEval Overall from $0.7927$ to $0.8093$, and beats every single-teacher and unfactored multi-teacher variant. In partial-noise reconstruction, independent panoptic, depth, and surface-normal probes show up to 53.1% PQ gain and 36.0% depth AbsRel reduction at $\tau=0.5$ across COCO and Flickr30K. From random initialization, PixelDense also reaches the baseline's peak GenEval $1.23\times$ faster.