NeurIPS 2026 论文列表

Vision-Language-Action (VLA) models inherit rich semantic priors from pretrained vision-language models, yet these priors do not guarantee persistent instruction sensitivity after narrow-domain fine-tuning. We observe a systematic failure mode: although VLA models remain visually reactive, their later actions can become weakly constrained by the commanded instruction, leading to unstable object interaction and visually plausible but instruction-inconsistent behavior. We trace this behavior to a collapse of instruction-conditioned representations across network depth and task progress. More fundamentally, this collapse is enabled by the learning objective itself: standard imitation learning admits Bayes-optimal solutions that are invariant to language. We address this by recasting instruction grounding as identifying the commanded instruction among counterfactual alternatives. Based on this view, we propose Counterfactual Instruction Grounding (CIG), a contrastive objective that encourages the generated trajectory to remain identifiable with the commanded instruction among counterfactual alternatives. CIG applies to both autoregressive and flow-matching VLA models, using exact action-chunk likelihoods for autoregressive models and an energy-based pseudo posterior for flow-matching models to avoid intractable trajectory likelihood estimation. Practically, CIG can be applied directly to already fine-tuned models as a continuation-stage grounding objective, restoring instruction sensitivity without retraining from scratch or changing the architecture. Extensive experiments on RoboCasa and LIBERO-Plus demonstrate that CIG achieves strong instruction-following performance with success rates of 71.3% and 86.3%, respectively. Real-world deployment further shows that CIG transfers to kitchen manipulation and preserves instruction-conditioned behavior in visually ambiguous scenes.

Counterfactual explanations for time-series classification generate synthetic instances that flip a prediction to a desired class while remaining plausible and making small, local changes. Existing methods often rely on classifier gradients, and many do not naturally extend to one-class settings. To bridge this gap, we propose CTCF, a model-agnostic method for hard-decision classifiers in supervised and one-class settings. CTCF decouples generation from classifier-specific control by reusing an unconditional Flow Matching generator and learning a flow-time-dependent desired-class region from interpolation states labeled by endpoint hard decisions, without classifier gradients or class probabilities. At inference time, CTCF uses a dual-control mechanism: fused-lasso endpoint-objective steering encourages small, segment-local changes, while linearized projection correction reduces violations of a conservative desired-class region. Theoretically, under ideal marginal matching, we show that the surrogate risk on interpolation states equals the corresponding ODE-trajectory risk. Experiments on UCR datasets demonstrate CTCF's effectiveness across quantitative metrics and qualitative case studies.

Estimating counterfactual outcomes when multiple treatments are applied simultaneously, the composite treatment setting, is fundamentally harder than the single-cause case: the treatment space grows combinatorially, most combinations are unobserved, and confounding is entangled across dimensions. We propose PACE (Progressive Alignment for Composite Effects), which exploits a known causal DAG over treatment variables to progressively transform the observational data distribution into the target interventional distribution. PACE performs K sequential augmentation steps in reverse topological order, each isolating a single treatment component and conditioning only on its parents in the DAG. We derive the closed-form optimal augmentation density minimizing KL divergence at each step, establish finite-sample bounds prescribing the augmentation budget via DAG-aware selection bias ratios, and show that the DAG provides a deterministic processing order that eliminates the need for random permutation averaging. On semi-synthetic benchmarks with real covariates (IHDP, Twins, Hillstrom) and controlled synthetic environments, PACE substantially outperforms seven baselines on both continuous and binary outcomes, with the largest gains on rare treatment combinations where confounding is most severe. Ablations confirm consistent improvements over the closest prior method across varying DAG topologies, confounding strengths, and under DAG misspecification.

Large language models are increasingly deployed as autonomous agents that plan and act over long-horizon interactions, learning from both final outcomes and rich environmental feedback. However, optimizing from such interaction experience remains challenging, as feedback is heterogeneous and rarely carries explicit reward information. Existing post-training approaches rely on outcome-driven methods, such as reinforcement learning with verifiable rewards (RLVR), which primarily exploit final success signals while leaving interaction experience and feedback underutilized. In sparse-reward, long-horizon tasks, this often results in **distribution sharpening**: the policy reinforces a narrow set of already-successful behaviors, without substantially improving the feedback-grounded agency needed for broader problem-solving capacity (e.g., Pass@$k$). We propose ***Counterfactual Distillation***, a framework that converts exploration-derived experience into trainable policy supervision. Specifically, we organize exploration as a tree-structured search process, where the agent reflects on past decision points and makes new experience-guided decisions to form alternative branches. These corrections are then distilled as counterfactual targets: we train the model to predict the revised action from the original history without explicitly providing the experience, thereby introducing out-of-distribution behavior patterns and expanding exploration capacity. Across diverse interactive coding and agentic tasks, our method outperforms outcome-driven baselines such as GRPO and experience-based methods such as Early Experience, with gains of up to **14%** on Pass@128. When interleaved with reinforcement learning updates, it further raises the performance ceiling, yielding over **10%** improvement in Pass@1. We provide an [anonymized implementation](https://anonymous.4open.science/r/Counterfactual-Distillation-0E71).

Policies achieving strong performance in simulators or learned world models can fail when deployed into the real environment. While there must be environmental differences that account for these failures, not all differences are equally responsible. A benign error of irrelevant visual details may co-exist with a failure to predict a single critical transition causing an inevitable catastrophic failure. In this paper, we introduce \emph{counterfactual debugging} as an approach for \emph{world model transfer gap attribution} to identify the time steps whose transition or reward errors are \emph{causally responsible} for a performance degradation exhibited in a real trajectory, rather than only visually or statistically different. We develop a scalable algorithm for computing these attributions that exploits the sparsity of causal errors by recursively divide-and-conquer, significantly reducing computational costs and achieving an exponential speedup. The resulting ranked attribution report can better explain the performance gap in the real trajectory. Experiments across several distinct environments with a variety of injected world-model failures (observation corruption, reward misprediction, and physics violations) demonstrate that counterfactual debugging correctly identifies the errors that are responsible for performance gap, providing theoretically grounded, actionable insights for model improvement.

Large-scale human demonstrations have been a key ingredient in training interactive agents, from robot policies to computer-use agents and game world models. But for multi-agent world modeling, the open data landscape is missing a crucial combination: expert humans acting together in the same evolving world, observed from every participant’s viewpoint, with precise controls and state. Internet-scale video captures human behavior, but its controls are inferred and its viewpoints are not synchronized; simulator and policy rollouts provide exact labels and all-player views, but not strategic, collaborative human play. We present CounterStrike-1K, a public benchmark dataset built from professional Counter-Strike 2 match replay files. It contains 1,490 rendered player-view hours across seven maps, with every released round synchronized across all ten active player viewpoints. Each view is paired with audio, replay-derived game controls, player state, world position, and sparse game event annotations. Because each round records expert human teams coordinating under partial observability, CounterStrike-1K tests whether models can simulate not just game physics or single-view visual plausibility, but skilled decisions and their consequences across all players’ views. We provide fixed splits and benchmarks for action-conditioned prediction, cross-POV retrieval, multi-POV action coverage, and shared-state recovery. By releasing round-complete multi-perspective demonstrations at this scale, CounterStrike-1K provides a grounded training and evaluation testbed for multi-agent world-model research.

Counterfactual explanations (CFs) provide human-interpretable insights into model's predictions by identifying minimal changes to input features that would alter the model's output. However, existing methods struggle to generate multiple high-quality explanations that (1) affect only a small portion of the features, (2) can be applied to tabular data with heterogeneous features, and (3) are consistent with the user-defined constraints. We proposeCounterFlowNet, a generative approach that formulates CF generation as sequential feature modification using conditional Generative Flow Networks (GFlowNet). CounterFlowNet is trained to sample CFs proportionally to a user-specified reward function that can encode key CF desiderata: validity, sparsity, proximity and plausibility, encouraging high-quality explanations. The sequential formulation yields highly sparse edits, while a unified action space seamlessly supports continuous and categorical features. Moreover, actionability constraints, such as immutability and monotonicity of features, can be enforced at inference time via action masking, without retraining. Experiments on eight datasets under two evaluation protocols demonstrate that CounterFlowNet achieves superior trade-offs between validity, sparsity, plausibility, and diversity with full satisfaction of the given constraints.

We consider the problem of Cost-Aware Learning, where sampling different components of a finite-sum objective incurs different costs. The objective is to reach a target error while minimizing the total cost. We propose Cost-Aware SGD, which uses a distribution based on gradient norms and costs to sample components. We provide a thorough analysis of this algorithm, including cost-improvement bounds over baselines, a characterization of distribution proxy sub-optimality, and a lower bound. We apply our theoretical insights to reinforcement learning with language models, where the computational cost of sequence-level policy gradients varies with length. We find that the advantage magnitude serves as a high-fidelity proxy for gradient norms, and use this to introduce Cost-Aware GRPO. Empirical results on 1.5B, 4B, and 8B LLMs demonstrate that this algorithm significantly reduces the tokens used in policy optimization while matching or exceeding baseline accuracy.

We study the problem of auditing the fairness of a given classifier under partial feedback, where true labels are available only for positively classified individuals (e.g., loan repayment outcomes are observed only for approved applicants). We introduce a novel cost model for acquiring additional labeled data, designed to more accurately reflect real-world costs such as credit assessment, loan processing, and potential defaults. Our goal is to find optimal fairness audit algorithms that are more cost-effective than random exploration and natural baselines. In our work, we consider two audit settings: a black-box model with no assumptions on the data distribution, and a mixture model, where features and true labels follow a mixture of exponential family distributions. In the black-box setting, we propose a near-optimal auditing algorithm and show that a natural baseline can be strictly suboptimal. In the mixture model case, we design a novel algorithm that achieves significantly lower audit cost than the black-box case under mild assumptions. Our approach develops novel connections between partial feedback and learning from truncated samples. We also extend known results on maximum-a-posteriori oracles for spherical Gaussian mixtures to handle exponential family mixtures, which may be of independent interest. Moreover, our algorithms apply to popular fairness metrics including demographic parity, equal opportunity, and equalized odds. Empirically, we demonstrate strong performance of our algorithms on real-world fair classification datasets like Adult Income and Law School, consistently outperforming natural baselines by around 50% in terms of audit cost.

NIGHTS is a recent benchmark for human perceptual similarity over objects, scenes, and composition. DreamSim, the supervised metric introduced with NIGHTS, and later supervised metrics trained on the same data report large gains over self-supervised baselines. The standard NIGHTS evaluation uses the unanimous-agreement subset of a released 100,000-triplet corpus, leaving weak-majority and split-vote triplets outside the headline benchmark. We evaluate eleven perceptual similarity methods on the full released corpus, analyzing unanimous, weak-majority, and split-vote triplets separately. On the 64,420 non-tie triplets, cosine distance on a frozen self-supervised backbone reaches 74.1% accuracy without training, LoRA fine-tuning reaches 77–79%, and the DreamSim ensemble reaches 80.7%. A split-half reproducibility estimate from the triplet vote counts shows that majority labels reproduce only 70.1% of the time overall, and only 33–50% on weak-majority triplets. The extra accuracy comes from matching the collected votes, not from showing that the metric would better match a new set of human judgments. The evaluation target is also unstable across task formulations: NIGHTS’ 2AFC similarity labels and JND same/different labels disagree on 51–64% of shared stimuli. Fine-tuning changes the feature basis as well, increasing foreground-only fragility and making DreamSim the most fragile method under rotation, color, and blur. NIGHTS supports a narrower conclusion: cosine distance on self-supervised features captures the stable signal currently measurable in these labels. Additional claims of human alignment require a redesigned benchmark with more votes per triplet, axis-specific similarity questions, per-subject statistics, and ablation tests.

Decoding cognitive functions and brain diseases from functional magnetic resonance imaging (fMRI) is a foundational problem in neuroscience advanced by recent deep learning across task and resting-state (rs) settings. Existing region-level decoders capture spatial structure through region-of-interest (ROI)-aware graph or Transformer architectures, but apply a uniform temporal operator across all regions. Consequently, they do not explicitly model how cortical regions integrate information over heterogeneous timescales. We propose CoRTeX, a region-resolved recurrent architecture for fMRI in which each cortical region carries its own learnable parameters and evolves on its own learned timescale. At its core, per-region differentiable temporal integration windows allow each ROI to learn how far back in time to integrate information through a soft mask over a private buffer. Cross-region information is restricted to a single identity-aware attention channel, additionally biased by a co-activation-driven memory whose learning and forgetting rates are themselves region-specific. Experiments on stimulus classification using task-fMRI (NSD) and on two disease diagnosis tasks using rs-fMRI (ABIDE, COBRE) demonstrate that CoRTeX achieves competitive performance with strong baselines while enabling region-level analyses that are not well-defined for models with exchangeable units: the learned per-region timescales exhibit interpretable spatial organization, and class-specific regional attribution aligns with category-selective cortex. Code is available at: https://anonymous.4open.science/r/CoRTeX-F.

Chain-of-thought (CoT) monitoring is increasingly relied upon to detect misbehavior in LLMs, under the implicit assumption that misbehaving actors must reason about their misbehavior, leaving detectable traces in the CoT. We show this assumption fails for a class of attacks we call planning-execution decoupled attacks, in which misbehavior is injected via a corrupted plan before the actor begins reasoning. Using the investigator agent framework of Li et al. [2025], we discover that exposing an actor to a corrupted reasoning plan upstream causes it to reproduce the flawed reasoning in its own CoT, embedding misbehavior into natural-looking reasoning with few suspicious traces. The attack scales to stronger reasoning models and harder tasks, steering them toward misbehavior while evading detection. Evaluating a suite of thinking and non-thinking monitors, we find that thinking monitors substantially outperform non-thinking ones, though even the best miss a meaningful fraction of attacks. Critically, the relationship between monitor thinking budget and detection is not monotonic: extra reasoning sometimes improves detection but can also hurt it when monitors talk themselves into accepting corrupted CoTs as benign. This refines Guan et al. [2025], who show thinking budget generally helps monitoring; we find detection also depends on whether the additional reasoning is directed toward critical evaluation rather than rationalization.

Network-level traffic signal control (TSC) involves coordination under partial observability and strict real-time constraints. In real-world urban networks with dozens or hundreds of intersections, direct coordination incurs an intractably large joint action space, making centralized cooperation difficult to deploy in practice. We propose CorridorLight, which tackles this challenge by casting cooperation as task negotiation. A high-level Corridor Agent represents the network as a direction-level graph and periodically generates a small set of junction-disjoint corridor tasks via a seed-and-grow procedure. These tasks provide lane-level preference guidance, while Intersection Agents maintain fast decentralized control and selectively support the active corridor task. By learning when to cooperate versus act selfishly, CorridorLight enables an adaptive trade-off between local delay minimization and corridor-level relief, leading to improved global performance. To this end, we derive a closed-form, Lipschitz-continuous gating function by maximizing a counterfactual net gain between cooperative and selfish behavior, yielding smooth, interpretable cooperation without brittle heuristics. Experiments on synthetic grids and real-world networks show consistent system-level improvements and outperform cooperative TSC baselines across diverse topologies and demand patterns.

Correspondence pruning—separating sparse inliers from severe outliers in noisy feature matches—is a long-standing bottleneck for two-view geometry estimation in 3D vision. Existing learning-based pruners pursue this goal through ever more sophisticated context operators—local graphs, dynamic neighborhoods, or global attention—yet a controlled comparison reveals a striking phenomenon: when we freeze the structural assignment of these operators after a single forward pass, accuracy collapses by over $30\%$ mAP regardless of operator choice. This suggests that the limiting factor of current methods is not which operator is used, but the implicit assumption that structure can be committed in one shot. We argue that correspondence pruning should instead be cast as \emph{iterative structural rectification} (ISR), in which the geometric structure and the feature representation must \emph{co-evolve} layer by layer. Building on this view, we propose \textbf{ISR-Net}, a minimal architecture that pairs a layer-wise re-derived local-context operator (a dynamic $k$-NN graph in feature space) with a layer-wise re-derived global-context operator (soft assignment to learnable cluster prototypes), so that any gain over prior arts is attributable to the rectification \emph{schedule} rather than to operator novelty. Despite using deliberately off-the-shelf operators, ISR-Net attains state-of-the-art results on YFCC100M, SUN3D, and HPatches, surpassing the strongest attention-based competitor by $+8.2\%$ mAP@$5^\circ$ while running $35\%$ faster, indicating that \emph{how often structure is refined} matters more than the operator's complexity.

Existing methods for Virtual Try-On (VTON) often struggle to accurately transfer garment appearance, especially in unpaired settings where accurate person-to-garment correspondence is required. These methods do not explicitly supervise person-to-garment alignment, leaving correspondence to be learned implicitly within the generation model. In this paper, we first analyze full self-attention in DiT-based architectures and show that person-to-garment query-key matching aligned with local geometric correspondence is closely related to try-on quality. Building on this insight, we introduce CORrespondence ALignment (CORAL), a DiT-based framework that explicitly aligns query-key matching with robust external correspondences. CORAL integrates two complementary components: a correspondence distillation loss that aligns reliable matches with person-to-garment attention, and an entropy minimization loss that sharpens the attention distribution. For evaluation, we further propose a VLM-based evaluation protocol to better reflect human preference. CORAL consistently improves over the baseline, enhancing both global shape transfer and local detail preservation. Extensive ablations validate our design choices. Code and weights will be publicly available.

We study contextual bandit problems with correlated arms and access to surrogate reward signals produced by a machine learning model, motivated by applications such as large language model (LLM) routing. Unlike classical contextual bandits that rely solely on bandit feedback and assume conditional independence across arms, our setting allows context-dependent inter-arm correlations and auxiliary reward information that may be noisy or misspecified. We propose algorithms that leverage such surrogate rewards through two complementary designs. A coupled reward-mixing approach pools true and surrogate rewards to accelerate learning when surrogate signals are reliable, while a decoupled prediction-mixing approach maintains separate estimators for bandit feedback and surrogate rewards and adaptively combines their predictions. This decoupling yields robustness to surrogate misspecification, recovering regret guarantees comparable to reward-only bandit methods in the worst case, while achieving improved regret when surrogate predictions are sufficiently informative. We provide theoretical regret analyses for both approaches and evaluate them on LLM routing benchmarks under varying accuracy versus cost trade-offs. The results demonstrate improved sample efficiency and consistently better accuracy–cost trade-offs compared to standard contextual bandit baselines and strong static routing methods.

While Diffusion Language Models (DLMs) are theoretically well-suited for iterative refinement due to their non-causal structure, off-the-shelf masked diffusion language models (MDLMs) often fail to reliably revise incorrect tokens in practice. The key challenge lies in the model's inability to distinguish between correct and erroneous tokens in a visible sequence. Standard MDLM training is restricted to the objective of unmasking, undermining the effectiveness of refinement guided by confidence. Based on this observation, we study corrective behavior in DLMs, defined as the ability to assign lower confidence to incorrect tokens and iteratively refine them while preserving correct content. Under matched compute and data, we show that this capability is not induced by conventional masked diffusion objectives and propose a post-training principle oriented by correction that explicitly supervises visible incorrect tokens, enabling discriminative confidence and targeted refinement. To evaluate corrective behavior, we introduce the Code Revision Benchmark, a controllable and executable benchmark for assessing error localization and in-place correction. Beyond improving in-place correction, our post-training principle yields stronger general generation as a direct consequence: the same error-aware confidence that drives reliable revision also makes parallel decoding more effective and improves answer quality on standard generation tasks. We validate this dual benefit across three distinct domains, code completion, mathematical reasoning, and high-entropy parallel decoding (ParallelBench), where models trained with our objective consistently outperform both the base model and standard masked-diffusion fine-tuning.

In tool-integrated reinforcement learning with verifiable rewards (RLVR), sequence-level verifier signals are issued for trajectories whose outcomes are determined by a sparse subset of tokens—tool names, argument keys, and argument values. Standard RLVR assigns the same advantage to every token in a rollout, diluting credit at precisely the positions that matter. We propose **Correctable Fork Tokens (CFT)**, which decomposes selective credit assignment into two subproblems: *where* to update, and *how much* to update at those positions. To this end, CFT introduces a training-only answer-conditioned branch sharing parameters with the student policy. The key insight is that conditioning on the ground-truth answer concentrates the model's distribution at decision-fork positions, providing hindsight localization without additional parameter capacity or imitation targets. Critically, this benefit requires accurate answer information: a larger-capacity teacher without the answer fails to replicate these gains, confirming that privileged answer information—not teacher capacity—is the active ingredient. CFT uses this entropy reduction ($IG_t$) to identify correctable fork positions, then rescales token-level credit at those positions while preserving the verifier advantage as the sole signal for update direction. On BFCL v3/v4, CFT improves multi-turn accuracy by **4.2/6.5** pp and overall accuracy by **3.1/2.2** pp over GRPO; it further surpasses On-policy distillation—which employs a larger Expert Model as teacher—by **3.6/6.7** pp on multi-turn, confirming that answer-conditioned fork localization rather than teacher capacity drives the improvement. On Tau3, the average score improves by **8.3** pp. We additionally present a three-level structured verifier for tool-call settings; ablations confirm that finer-grained verifier feedback and selective token credit are complementary. Our reproducible implementation is available at [https://anonymous.4open.science/r/CFT](https://anonymous.4open.science/r/CFT).

Human labels are important supervisory signals for preference-based reinforcement learning, but their collection is often constrained by practical considerations that standard training objectives tend to ignore. For instance, Direct Preference Optimization (DPO) usually treats observed preference labels as clean, even though real annotation pipelines rely on heterogeneous annotators whose reliability varies across individuals and with the difficulty of each comparison. This setting introduces three challenges: (1) Noisy labels introduce systematic bias into the preference optimization objective and need to be rigorously corrected. (2) The allocation of pairwise comparisons to different annotators is a design problem that directly affects the statistical efficiency of the training pipeline. (3) Non-random routing creates selection bias in annotator evaluation, complicating downstream decisions such as compensation. We address these challenges within a unified framework for DPO that combines annotator-aware posterior correction with information-guided annotator routing. Posterior correction leverages an instance-dependent noise model to infer latent ground-truth preferences from noisy labels, mitigating bias in the DPO objective. Annotator routing is formulated as an optimal experimental design problem that allocates comparisons to annotators according to their expected statistical information under a fixed annotation budget. To support reliable annotator evaluation under non-random routing, we introduce a doubly robust estimator that corrects for selection bias. Semi-synthetic experiments on the UltraFeedback dataset show that our approach improves ground-truth preference recovery and downstream AlpacaEval 2 performance over baselines, with consistent gains across multiple open-source LLM model architectures. Annotator reliability-estimation experiments further show that the doubly robust estimator substantially reduces mean squared error relative to other baselines, demonstrating its value for annotator evaluation.