NeurIPS 2026 论文列表

Sharpness-Aware Minimization (SAM) is widely used in domain generalization to promote flat minima that generalize to unseen distributions. Existing SAM-based methods typically penalize the sharpness of the *average* loss across domains. We show that this prevailing formulation—termed AvSAM—has a fundamental geometric limitation: it can favor solutions with misaligned domain-wise dominant Hessian eigenvectors. As a result, such solutions may appear flat on average while remaining sharp within individual domains. To address this limitation, we propose an alternative objective that penalizes the average *domain-wise* sharpness. We show that this objective decomposes into the AvSAM objective plus a non-negative residual term, which we call the *Cancellation Gap*. This gap is minimized if and only if the domain-wise dominant eigenvectors are aligned up to sign, indicating that it mitigates AvSAM's tendency to favor misaligned solutions. However, we further uncover a degeneracy: the Cancellation Gap can become small without improving alignment when sharpness is highly imbalanced across domains. To mitigate this failure mode, we propose Cancellation Gap Minimization (CGM), which augments the objective with a squared coefficient-of-variation ($\mathrm{CV}^2$) regularizer to discourage sharpness imbalance. Experiments show that CGM achieves the best average performance across standard domain generalization benchmarks among strong baselines. Diagnostic analyses further confirm reduced domain-wise sharpness and improved alignment of dominant Hessian eigenvectors.

Vision-language-action (VLA) models have achieved strong performance in embodied manipulation, but still lack a clear mechanism to balance behavioral stability with task-semantic sensitivity. We identify two complementary failure modes. Under task-preserving changes, where task semantics remain unchanged but scene appearance varies (e.g., style, illumination, clutter, or paraphrasing), policies often exhibit unnecessary action drift. Conversely, under semantic-breaking changes, where key task semantics such as the target object or constraint are altered, policies frequently fail to produce sufficiently distinct behaviors and instead follow the original trajectory. To address this gap, we propose BAS-VLA, a task-semantic action calibration framework built on top of a frozen base VLA. BAS-VLA adopts a breaking-centered calibration core as the default path, and introduces a selective evidence-gated preserving auxiliary that activates only when nuisance variation is detected while task semantics remain consistent. On the OpenPI-pi0.5 / LIBERO-Object Milk-Swap benchmark, BAS-VLA maintains high success on clean (98.0%) and semantics-preserving conditions (97.5%), while reducing clean-criterion success to 0.0% under deliberate target-object swaps, demonstrating strong stale-task suppression and task-semantic separation. On validated style-preserving shifts, it improves success from 42% to 70% without degrading clean performance. These results highlight that reliable VLA behavior requires moving beyond appearance robustness toward explicit task-semantic action calibration.

Existing depth-compression methods for large language models (LLMs) operate on layers individually or restrict grouping to adjacent or contiguous positions, implicitly assuming functional redundancy is local. We challenge this assumption by computing full pairwise Centered Kernel Alignment (CKA) similarity matrices across thirteen LLMs spanning six architecture families and 1.6B--14B parameters, revealing that non-adjacent layers exhibit 3.6--15.9$\times$ more high-similarity pairs than adjacent ones. We propose Fusion via Community Mining (FCM), which builds the full $L \times L$ similarity graph, applies spectral community detection to recover clusters of functionally equivalent layers, retains the highest-Block-Influence layer per cluster, and applies LoRA recovery. At 2$\times$ compression, FCM lowers mean perplexity below ShortGPT on twelve of thirteen models, with statistically significant improvements on ten under a Welch's $t$-test with Holm--Bonferroni correction (up to 50\% on Qwen3-4B, $p_\text{corr}<0.001$); two rows are ties and Mistral-7B is the lone loss ($+$14.3\%). On three anchor models, FCM also outperforms three recent adjacent-pair methods (FlattenGPT, SWM, DeltaLLM) by 7.8--35.3\% under matched calibration and recovery. The perplexity advantage transfers downstream: on a five-benchmark zero-shot reasoning suite (ARC-Easy, ARC-Challenge, HellaSwag, WinoGrande, PIQA) over four anchor models, FCM matches or exceeds ShortGPT on 15 of 20 (model, benchmark) cells, and the 2$\times$-compressed Qwen2.5-7B realises 46\% fewer parameters, 42\% lower peak VRAM, and 84\% higher decode throughput on a single A100. Ablations confirm spectral grouping outperforms Fisher contiguous segmentation by 41\% and uniform partitioning by 51\%, and a compression sweep reveals a 2$\times$ crossover beyond which FCM's advantage grows monotonically and persists at the 14B scale. Code: \url{https://anonymous.4open.science/r/fcm-layer-fusion-D820}.

Flow matching has recently emerged as a promising alternative to diffusion-based generative models, particularly for text-to-image generation. Although flow matching places no restriction on the source distribution, most existing systems still inherit a standard Gaussian from diffusion models, and the source is rarely treated as an optimization target at this scale. Recent works have begun to revisit this choice through condition-dependent or learned sources, yet evidence that such designs are effective in modern text-to-image systems—with high-dimensional latents and tightly integrated conditioning—remains limited. In this work, we study condition-dependent source distributions for flow matching along three axes: _why_ source learning helps, through the lens of the intrinsic variance term in the flow-matching objective; _how_ to make it work in modern text-to-image systems, where variance-only regularization and directional source—target alignment are critical for stable end-to-end training; and _when_ it is most beneficial, by connecting source design to recent representational advances in generative modeling and identifying target representation regimes in which learning the source yields the largest gains. Extensive experiments across multiple text-to-image benchmarks, backbones, and scales demonstrate that principled source design yields consistent and robust improvements—including up to $\mathbf{3.01\times}$ faster convergence in FID and $\mathbf{2.48\times}$ in CLIP score—and outperforms representative prior conditioned-source and condition-aware coupling methods.

Language models (LMs) obtain much of their capability from the diversity of their training data. This is possible because neural architectures like transformers are remarkably flexible and can handle a variety of domains such as natural language, code, and mathematics. This raises the question: are transformers optimal for any specific domain? In this position paper, we present empirical evidence that the inductive biases of transformer-based LLMs should be re-examined, since, while broadly useful, they are often suboptimal on subsets of their training data. **Methods.** To support this position, we use a method that learns new activation functions within a transformer as a tool to explore the space of inductive biases. We train the resulting modified architectures on various data subsets. A large-scale evaluation shows that vanilla transformers are relatively easy to improve, but mostly on a per-domain basis. For example, an architecture optimized for a natural language dataset like FineWeb performs well on another such dataset like TinyStories, whereas architectures optimized for mathematics perform poorly on natural language. We also identify the mechanisms responsible for these improvements. **Implications.** These results call for re-evaluating core design choices in language models. We do not advocate a return to handcrafted, domain-specific models. Instead, we emphasize that different domains can benefit from different learning mechanisms. This will require new designs that preserve cross-domain interactions, analogous to how reasoning in the brain arises from the interaction of multiple specialized networks. We argue that this perspective may be key to improving reliability and data efficiency.

Diffusion models achieve strong generative performance but often struggle to align generated samples with human preferences measured by a reward model. A simple yet effective algorithm for test-time alignment is Best-of-$N$ (BoN) sampling, which draws $N$ i.i.d. samples from a pre-trained diffusion model and outputs the single highest-reward sample. Despite its empirical success, this procedure is inherently post hoc: reward information is used only for selection after generation, not for improving the reverse diffusion process itself. Consequently, BoN sampling does not improve the average alignment of generated samples and is primarily suited to single-output settings. We propose Best-of-$N$ Guidance (BoNG), a novel method that integrates the principle of BoN sampling directly into the reverse diffusion process. BoNG performs online BoN selection over denoising particles and adjusts the reverse diffusion process to steer the particle population toward higher-reward regions during generation. Specifically, by introducing an asymmetric guidance interaction among denoising particles, BoNG uses the current BoN particle as a guidance signal to the rest of the particle population. This particle-level interaction reshapes the sampling process toward higher-reward regions, enabling BoNG to improve not only the final best sample beyond Vanilla BoN sampling, but also the average quality of generated samples. Over 36 empirical comparisons, BoNG achieves the best performance in 29 cases, ranking first in 80.56\% of the comparisons against SMC and Vanilla BoN sampling. BoNG also supports multi-output capability, showing 1.3$\times$ ImageReward score than the latest sample-based guidance method with 1.6$\times$ speedup.

We study fixed-confidence best arm identification in generalized linear bandits under a hybrid feedback model: at each round, the learner may query either (i) absolute reward feedback from a single arm or (ii) relative (dueling) feedback from an arm pair, both governed by generalized linear models. We introduce a likelihood-ratio–based confidence sequence that unifies heterogeneous generalized linear observations and yields an explicit ellipsoidal confidence set under a self-concordance assumption. Building on this confidence set, we propose a hybrid Track-and-Stop algorithm that adaptively allocates queries by tracking a minimax-optimal design over a joint action space of arms and pairs. We establish $\delta$-correctness and provide high-probability upper bounds on the stopping time. We further extend the framework to a cost-aware setting that accounts for heterogeneous acquisition costs across feedback modalities. Empirical experiments demonstrate that the proposed algorithms significantly improve sample efficiency over baseline methods.

We study the best arm identification problem in a novel non-stationary environment that we coin Shifting Means. While classically the mean rewards of the $K$ arms are stable in time, in Shifting Means only the gaps $\boldsymbol{\Delta}$ between mean rewards are stable, while their common shift may be determined adversarially in each round. The objective of the learner is to identify the best arm with high probability while minimizing sample complexity (the fixed confidence setting). Handling shifts requires new tools: we show that algorithms employing a Generalized Likelihood Ratio Test (GLRT) stopping rule, including the popular Track-and-Stop, fail under time-varying shifts. Instead, we propose Importance Weights for Shifting Means ($\mathsf{ISM}$). Assuming means bounded by $U$ and $\sigma^2$-sub-Gaussian rewards, we show $\mathsf{ISM}$ to be $\delta$-correct and to enjoy a sample complexity bound of $K (\sigma^2 + U^2) \Delta_{\min}^{-2} \ln \frac{1}{\delta}$. We also present a matching (up to constant factors) worst-case lower bound and evaluate our results empirically.

Binary diffusion models typically require a massive number of function evaluations (NFEs) to generate high-quality samples, making practical inference computationally expensive. Reducing NFEs while preserving sample quality without relying on distillation or additional training remains a significant challenge. However, existing binary diffusion models sequentially define a discrete one-step forward path and subsequently derive the reverse posterior. In low-NFE scenarios that require cross-step sampling, these models incorrectly approximate the true multi-step likelihood using a single-step likelihood formulation, which severely degrades sample quality. To address this fundamental limitation and completely decouple the generative dynamics from fixed discrete time steps, we propose Bernoulli Flow Models (BFM). Rather than building upon sequential one-step Markov diffusion chains, we predefine a unified continuous global Bernoulli probability flow path between data distributions and pure noise, from which we derive analytical closed-form posterior transitions over arbitrary time intervals. Consequently, reducing the inference NFE in BFM is no longer an approximation of skipped discrete steps; it simply requires re-evaluating the analytical posterior on the new time intervals. This eliminates the structural training-inference mismatch inherent to discrete chains, yielding strictly self-consistent low-NFE sampling. Experimental results show that BFM is highly robust to aggressive NFE reduction. On the LSUN Churches 256x256 dataset, a 256-step-trained BFM achieves an FID of 9.31 when sampled with only 16 steps, whereas the state-of-the-art discrete baseline severely degrades to 204.10. BFM also remains competitive with both continuous and discrete generative baselines under standard full-step inference. Ultimately, these results establish BFM as a theoretically rigorous, self-consistent, and practically effective framework for fast binary data generation.

Multimodal large language models (MLLMs) and diffusion models have each reached remarkable maturity: MLLMs excel at reasoning over heterogeneous multimodal inputs with strong semantic grounding, while diffusion models synthesize images and videos with photorealistic fidelity. We argue that these two families can be unified through a simple division of labor: MLLMs perform semantic planning, while diffusion models render pixels from high-level semantic guidance and low-level visual features. Building on this idea, we propose \textbf{Bernini}, a unified framework for video generation and editing. An MLLM-based \textbf{planner} predicts the target semantic representation directly in the ViT embedding space, and a DiT-based \textbf{renderer} synthesizes pixels conditioned on this plan, augmented by text features and, for editing, source VAE features for detail preservation. Because semantics serve as the interface, the planner and renderer can be trained separately and only lightly co-trained, preserving the pretrained strengths of both components while keeping training efficient. To better handle multiple visual inputs, we introduce Segment-Aware 3D Rotary Positional Embedding (SA-3D RoPE), and further incorporate chain-of-thought reasoning in the planner to better transfer understanding into generation. Bernini achieves state-of-the-art performance across a wide range of video generation and editing benchmarks, with the MLLM's pretrained understanding translating into strong generalization on challenging editing tasks.

Asymptotic e-values are emerging as a powerful alternative to asymptotic p-values, particularly in post-hoc inference and multiple testing, where significance levels may be data-dependent. Existing asymptotic e-values, however, suffer from the ``missing factor,'' a scaling inefficiency resulting in overly conservative inference. Drawing on the framework of near-optimal concentration inequalities developed by Bentkus in the 2000s, we introduce Bentkus-type asymptotic e-values and prove that they successfully eliminate the missing factor. We also demonstrate both theoretically and empirically that Bentkus-type e-values consistently deliver sharper inference than existing alternatives, leading to tighter post-hoc confidence intervals and higher rejection rates in multiple testing procedures.

Reinforcement learning (RL) is now standard for post-training large language models. The same reward optimization that elicits useful capabilities, however, can also reinforce latent backdoors planted earlier in the pipeline: attack success rates under 0.2% after SFT rise to 40--98% on training-distribution inputs after RL, and generalize to 13--30\% on held-out evaluation, with no modification to the RL data, reward function, or training loop. We study this phenomenon in the context of agent models, where SFT poisoning teaches the victim model (Qwen3-8B and a small frontier model) to call an attacker-controlled oracle. Because the oracle can be set up to always return correct answers, calling it earns higher reward than the model's own attempts, and RL reinforces the behavior. We further show that the oracle's RL-time responses can instill persistent biases, such as brand preferences, that survive into deployment even when no tool calls happen. Such patterns appear difficult to detect with current tools: traditional guard models do not flag this mechanism, and LLM-based auditors remain unreliable, achieving only 4.5\% precision even after iterative prompt refinement. Our findings point to the value of post-RL safety evaluation, in particular scrutinizing tool-call patterns such as unnecessary external invocations.

AI progress is increasingly tracked through evaluations, such as AI benchmarks, yet small perturbations in evaluation pipelines can produce unstable scores and model rankings. We recast benchmark reliability as a measurement problem and formalize it as a signal-to-noise ratio derived from a crossed random-effects decomposition. This framework yields a direct mapping between variance components and expected rank stability, linking theoretical reliability to observable leaderboard concordance. Modern AI benchmarks operate in a high-dimensional regime with many items and relatively few evaluated models, where classical item-level reliability measures are ill-posed. We define a lower bound for benchmark item-level reliability classically, and we further introduce a tractable proxy, $\lambda^\bigstar_6$, that preserves the variance structure underlying reliability without modifying the estimand through sparsity or dimensionality reduction. Across diverse benchmarks, we show that selecting items via $\lambda^\bigstar_6$ produces smaller subsets that achieve higher ranking reliability than random subsampling at fixed evaluation budgets. Finally, we analyze reliability under positive distribution shift--as often observed in the current AI ecosystem--by partitioning models into lower- and higher-performing cohorts. We show that reliability is population-dependent and frequently declines as between-model variance shifts among stronger systems. Our results establish a principled framework for diagnosing and improving benchmark reliability and demonstrate that ranking stability is neither intrinsic nor static, but a measurable and optimizable property of the benchmark–population pairing.

Biosynthetic gene clusters (BGCs) are co-located genes that encode the biosynthesis of secondary metabolites, a major source of clinically used antibiotics and anticancer agents. Recent advances in deep learning, particularly self-supervised foundation models, have spurred growing interest in BGC sequence modeling, but evaluation infrastructure has lagged behind. Current practice often relies on coarse classification accuracy over small, experimentally curated datasets, making it difficult to discriminate model capabilities or assess downstream utility. We introduce BGC-Bench, an evaluation suite designed to resolve this limitation by systematically incorporating biochemical knowledge in the benchmark design. In terms of generalization, we test whether representations transfer to genetically and chemically novel samples; for functionality, we examine application-relevant tasks, including retrieval, prediction, and active learning. BGC-Bench reveals that no single model class dominates: pretrained foundation models perform strongly under out-of-domain generalization but remain sensitive to training conditions; compositional baselines are competitive and lead in cross-modal retrieval; and active-learning gains vary substantially across tasks, strategies, and model families. Beyond these BGC-specific findings, BGC-Bench highlights a broader principle: domain-informed benchmark design over existing datasets enables better evaluation in data-scarce scientific domains.

LLM benchmarks are often treated as coherent measurement units, yet they are heterogeneous collections of instances spanning diverse domains, skills, formats, and contexts. As a result, aggregate benchmark scores can conflate model capability with benchmark composition, obscuring what existing benchmarks actually cover. We introduce BenchHub, a composition-aware evaluation framework that represents benchmarks as distributions over instance-level attributes. BenchHub integrates 54 benchmarks comprising 839K samples across 10 languages. It enables researchers to inspect benchmark contents, uncover reusable coverage hidden behind benchmark names, transparently compare benchmarks, and construct controllable evaluation sets for application-aligned model selection. Using BenchHub, we show that similarly motivated benchmarks can differ substantially in internal composition, existing benchmarks often contain reusable coverage beyond their stated purposes, models exhibit fine-grained category-level performance variation hidden by aggregate scores, and model rankings can shift under different reweighting and resampling configurations. Our results motivate evaluation practices that make benchmark composition explicit, inspectable, and controllable.

Structured generation remains a critical bottleneck for Multimodal Large Language Models (MLLMs). Current post-training paradigms struggle to resolve this: Supervised Fine-Tuning (SFT) hits a rigid ceiling, while scalar-reward Reinforcement Learning (e.g., DPO, GRPO, SimPO) suffers from a policy-level signal gap. Compressing long routing trajectories into sparse signals induces a short-completion bias and traps RL near the SFT baseline, whereas vanilla self-distillation actively regresses. To systematically quantify this joint visual-structural-relational reasoning, we introduce OracleGraph, a 10,000-page 2.5D historical document benchmark featuring a 12-dimensional diagnostic VQA suite and a structured-hallucination taxonomy. To overcome these optimization pathologies, we propose PRISM, a Policy-Reward Integrated Self-distillation framework. PRISM leverages multimodal information asymmetry via a Hindsight Teacher guided by a Diagnostic Graph Report (DGR). By synergizing Softmax Reward-Weighted Aggregation (preserving Graph-F1 rankings) with a Length-Normalized Auxiliary Preference Loss, PRISM explicitly neutralizes catastrophic truncation, yielding a $+3.81$pp Graph Reward gain and a $+2.3$pp schema-parsing improvement over SDPO ($p < 0.001$). While PRISM's nominal $+1.02$pp mean gain over SFT is not statistically significant ($p = 0.12$), it drastically stabilizes optimization, achieving $2.4\times$ lower cross-seed variance. Crucially, on a cross-book out-of-distribution (OOD) probe, PRISM's stability advantage amplifies to a $14\times$ variance reduction over SFT, confirming robust structural stabilization. Code and datasets are available at https://anonymous.4open.science/r/PRISM2026.

Many safety and alignment failures of large language models (LLMs) occur due to out-of-distribution (OOD) situations: unusual prompt or response patterns that are unforeseen by model developers. We systematically study whether LLM monitoring pipelines can detect these **OOD alignment failures** by introducing a benchmark called Misalignment Out of Distribution Benchmark (**MOOD**). It is difficult to find failures that are truly OOD for off-the-shelf models trained on vast safety datasets. We sidestep this by including a restricted training set in MOOD that we use to train our own monitors, as well as seven test sets with diverse alignment failures that are outside the training distribution. Using MOOD, we find that guard models (safety classifiers) often fail to generalize OOD. To fix this, we propose combining guard models with OOD detectors. We test four types of OOD detectors and find that a combination of a guard model and Mahalanobis distance-based OOD detector can improve recall from 39% to 47%. We also establish positive scaling trends across model scales for monitors that combine a guard model and OOD detector; we find that incorporating OOD detection into monitoring achieves a higher recall gain than using a guard model with 20 times more parameters. Our work suggests that OOD detection should be a crucial component of LLM monitoring and provides a foundation for further work on this important problem.

The rapid advancement of large language models (LLMs) has expanded their potential in the legal domain. However, existing legal benchmarks remain largely English-centric and oriented toward common law, leaving a critical gap in evaluating LLMs for civil law systems that govern most jurisdictions worldwide. To address this gap, we introduce Vietnamese Legal Benchmark VietLegal, a cognitively grounded benchmark designed for the hierarchical and codified structure of Vietnamese law. Although instantiated in Vietnamese legislation, VietLegal provides a replicable evaluation framework for civil law systems characterized by complex statutory hierarchies and frequent amendments. Inspired by Bloom’s taxonomy, VietLegal assesses multiple levels of legal understanding through tasks that mirror real-world legal assistant use cases, including legal question answering, multi-step reasoning, and scenario-based problem solving. The benchmark contains 10,450 expert-annotated samples, each cross-validated against authoritative legal sources to ensure fidelity to practical legal workflows. By offering the first standardized legal benchmark for Vietnamese, VietLegal enables systematic assessment of LLMs in civil law contexts and supports the development of more reliable and interpretable AI-assisted legal systems.

Cyber-physical system (CPS) forecasting models depend on sensor streams with noisy, biased, missing, or temporally misaligned readings, yet standard forecasting evaluation often selects models by nominal error without showing whether they remain robust under such faults. We introduce SensorFault-Bench, a shared CPS-grounded sensor-fault stress-test protocol for evaluating forecasting architectures and robustness-improvement methods, together with an operational taxonomy that organizes the method comparison. Across four real-world datasets and eight scored scenarios governed by a standardized severity model, it reports worst-scenario degradation, clean mean squared error (MSE), and worst-scenario fault-time MSE, separating relative robustness from absolute error. A disjoint fault-transfer split lets explicit fault-training methods train on adjacent fault families while reported scores use separate benchmark scenarios. Empirically, forecasting architectures favored by clean MSE can degrade sharply under faults, and clean-MSE rankings can disagree with worst-scenario fault-time error rankings. Chronos-2, the evaluated zero-shot foundation-model representative, matches or trails the last-value naive forecaster in clean MSE on the two single-target datasets and has the largest worst-scenario degradation on ETTh1 and Traffic, where all channels are forecast targets. For the evaluated robustness-improvement method set, paired deltas show selective degradation reductions: projected gradient descent adversarial training and randomized training lead where value faults dominate observed degradation, while fault augmentation leads where availability faults dominate. SensorFault-Bench provides open-source code, documented data access, and reproduction and extension guides, so new datasets, architectures, and robustness-improvement methods can be evaluated under the same CPS sensor-fault robustness protocol.