NeurIPS 2026 论文列表

On-policy distillation (OPD) is a promising approach for transferring reasoning capabilities to capacity-constrained models, yet sampled-token OPD often suffers from entropy collapse and negative transfer. We identify uniform credit allocation as a key bottleneck: existing objectives obtain dense teacher feedback on student-generated trajectories, but apply it uniformly across sampled tokens despite highly heterogeneous token-level signals. Most tokens carry redundant near-zero credit, while rare heavy-tailed negative credits can dominate updates and prematurely suppress plausible reasoning trajectories. We introduce REOPOLD (Relaxed On-Policy Distillation), a framework that relaxes this uniform assignment by controlling where teacher feedback is applied, how strongly it affects the update, and when the allocation rule changes during training. Across diverse reasoning tasks, REOPOLD improves sampled-token OPD over recent post-training baselines with up to 12$\times$ higher sample efficiency, and extends to cross-vocabulary distillation, self-distillation, and test-time scaling.

Reinforcement Learning has significantly advanced the reasoning capabilities of Multimodal Large Language Models (MLLMs), yet the resulting policies remain brittle against real-world visual degradations such as blur, compression artifacts, and low-resolution scans. Prior robustness techniques from vision and deep RL rely on static data augmentation or value-based regularization, neither of which transfers cleanly to critic-free RL fine-tuning of autoregressive MLLMs. Reinforcing reasoning against such corruptions is non-trivial: naively injecting degraded views during rollout induces reward poisoning, where perceptual occlusions trigger hallucinated trajectories and destabilize optimization. We propose ROMA, an RL fine-tuning framework that modifies the optimization dynamics to reinforce reasoning against visual degradation while preserving clean-input performance. A dual-forward-pass strategy uses teacher forcing to evaluate corrupted views against clean-image trajectories, avoiding new rollouts on degraded inputs. For distributional consistency, we apply a token-level surrogate KL penalty against the worst-case augmentation; to prevent policy collapse under regularization, an auxiliary policy gradient loss anchored to clean-image advantages preserves a reliable reward signal; and to avoid systematically incorrect invariance, correctness-conditioned regularization restricts enforcement to successful trajectories. On Qwen3-VL 4B/8B across seven multimodal reasoning benchmarks, our method improves robustness by +2.4% on seen and +2.3% on unseen corruptions over GRPO while matching clean accuracy.

Direct Preference Optimization (DPO) has recently been applied to medical image segmentation, since comparing candidate masks is substantially cheaper than producing new dense pixel-level annotations, making pairwise preference signals an appealing source of supervision. However, existing work derives preferences from ground-truth masks, leaving open whether the approach is sound under the imperfect feedback available in practice. In our analysis of this reformulation, we identify that candidate masks tend to agree over most of the image and differ only in localized regions, yet standard DPO averages the preference signal over the full spatial domain. This couples update strength to disagreement area, diluting small correct refinements and amplifying large misranked differences. Crucially, this bias even degrades performance even under oracle preferences derived from ground truth. We propose Region-Normalized DPO (RN-DPO), which normalizes the likelihood ratio over the disagreement region between candidate masks, removing this coupling. We further provide a systematic empirical study of preference-based segmentation fine-tuning under controlled noisy judges, analyzing the effects of mining strategies, judge types, and reliability regimes. Across two medical segmentation benchmarks, multiple judge configurations, and two backbone architectures, RN-DPO consistently improves over vanilla DPO and robust DPO variants, with gains persisting under noiseless oracle preferences.

Neural Operators (NOs) are powerful architectures for learning mappings between function spaces. While most advances focus on refining kernel parameterizations over the $d$-dimensional physical domain, the evolution of lifted embeddings remains underexplored, which often drives models toward computationally expensive embedding-scaling designs to improve approximation. In this paper, we introduce an auxiliary function dimension that models embedding evolution in operator form, thereby reformulating the NO pipeline in $d+1$ dimensions. We instantiate this framework via Fourier-based operators acting jointly on the physical and auxiliary domains, yielding a basis-diversified auxiliary evolution module as an alternative to brute-force embedding scaling. Across more than ten increasingly challenging benchmarks, ranging from the 1D heat equation to the highly nonlinear 3D Rayleigh–Taylor instability, our model consistently achieves the lowest relative $L_2$ error among the evaluated baselines. Crucially, this advantage is empirically supported by (1) controlled budget-aware comparisons against scaled and ablated baselines; (2) robustness under mixed-resolution training and super-resolution inference; and (3) zero-shot generalization to unseen temporal regimes. In addition, we present a broader set of design choices for lifting and recovery operators, demonstrating their impact on our model’s predictive performance.

Image tokenization aims to compress visual signals into latent codes that remain effective for reconstruction and generation. Driven by the multi-scale nature of images, organizing latent along a coarse-to-fine refinement process emerges as a promising direction for visual tokenization. Existing scale-wise tokenizers often introduce this structure through token ordering or latent prefixes, but each target scale is typically decoded directly from its prefix, making different prefixes behave like separate reconstruction codes. This can lead to redundant encoding of global structure across scales and limits the effectiveness of compact token budgets. We propose RefineTok, a scale-wise continuous tokenizer that makes decoding itself progressive. RefineTok carries a visual state across scales and uses each latent prefix to refine the previous state into the next scale, turning scale-wise latents into incremental refinement signals rather than independent reconstruction codes. This design yields a compact and structured latent interface for both reconstruction and downstream generation. Together with prefix-level training, a coarse-centric scale trajectory, and a semantic anchor, RefineTok achieves strong reconstruction fidelity among scale-wise tokenizers and competitive class-conditional generation on ImageNet-1K under compact latent budgets. Meanwhile, it preserves a progressive scale-wise interface that single-scale tokenizers do not provide.

Monocular 3D object detection spans two regimes: closed-set detectors operating within a fixed category vocabulary, and open-vocabulary detectors that localize arbitrary categories by leveraging depth foundation models for 3D geometry. We find that current depth foundation models, despite their strong zero-shot generalization, lack the object-level precision 3D detection demands: substituting a state-of-the-art depth foundation model for a strong detector's predicted depth degrades accuracy, even falling below the detector's own prediction. Rather than pushing detectors or depth models to be more accurate end-to-end, we treat object-level depth refinement as a stand-alone task and present RefineAny3D, a vision-language model that corrects depth without ever predicting a numerical value. Our key insight is that depth error has a direct visual signature in image space: when projected onto the image, a correctly placed box tightly encloses the object, while a too-far box projects too small and a too-close box projects too large. Depth refinement thus reduces to a visual alignment problem rather than a metric regression problem, which we instantiate by extending the VLM's vocabulary with action tokens that replace numerical depth output with categorical decisions, and by supervising the model on a large-scale chain-of-thought dataset that grounds each decision in explicit visual evidence. Applied as a single post-hoc step, RefineAny3D delivers consistent gains across closed-set detectors, open-vocabulary detectors, and 3D auto-labeling tools, and generalizes to novel categories, scenes, and cameras without retraining.

Coral reefs are rapidly declining under anthropogenic pressures (e.g., climate change), creating an urgent need for scalable and automated monitoring. Progress in data-driven coral analysis, however, is constrained by the scarcity of large-scale datasets with fine-grained labels that are taxonomically consistent across sites and studies. To address this gap, we introduce ReefNet, a large-scale public coral reef image dataset with point-level annotations mapped to the World Register of Marine Species (WoRMS) taxonomy. ReefNet aggregates imagery from 76 curated CoralNet sources and an additional reef site from Al-Wajh (Red Sea), totaling approximately 925K genus-level hard coral annotations. Through expert-driven verification and targeted filtering, we derive a high-confidence benchmark subset with 92% expert agreement over 39 hard-coral label classes, enabling reliable evaluation under realistic label noise and strong class imbalance. Beyond dataset construction, we establish a comprehensive benchmark spanning zero-shot, cross-domain few-shot adaptation, within-source evaluation, and cross-source transfer to the Al-Wajh dataset. Experiments with state-of-the-art vision–language models (VLMs), multimodal large language models (MLLMs), and vision-only backbones reveal substantial degradation in zero-shot and extremely few-shot regimes, while adaptation with in-domain supervision yields large gains yet still leaves a persistent gap under cross-source shift and on long-tail genera. These results highlight fundamental challenges in applying general-purpose multimodal models to biodiversity monitoring and underscore the importance of large-scale, taxonomically grounded, high-quality datasets. ReefNet serves as both a benchmark and a training resource for advancing fine-grained coral reef understanding. ReefNet data and trained models are provided on Hugging Face, and code for accessing and benchmarking the data is provided on GitHub.

Machine learning models deployed in decision-making pipelines, from vehicle routing to dynamic pricing to clinical decision support, must remain reliable in the face of low-quality training data. We study a predict-then-optimize setting in which an upstream model predicts the cost vector of a downstream linear program, with a large pool of noisy training data supplemented by a small curated set of clean data. We introduce Reduced Cost Influence Functions (RCIFs), which combine influence functions from robust statistics with the geometry of linear optimization to estimate how each training point shifts downstream decisions. Building on this, we propose a reweighting algorithm that uses RCIFs to downweight harmful training points and reduce downstream regret. Experiments on random linear programs and shortest-path problems show that our approach reduces regret, and accompanying theory characterizes properties of RCIFs and identifies regimes favorable to our method relative to baselines.

We study allowing large language models (LLMs) to process arbitrarily long prompts through the lens of inference-time scaling. We propose **Recursive Language Models** (**RLM**s), a general inference paradigm that treats long prompts as part of an external environment and allows the LLM to programmatically examine, decompose, and recursively call itself over snippets of the prompt. We find that RLMs can successfully process inputs more than an order of magnitude beyond model context window limits and, even for shorter prompts, dramatically outperform the quality of vanilla frontier LLMs and common long-context scaffolds (e.g., on GPT-5 by a median across the evaluated benchmarks of 26% against compaction, 130% against CodeAct with sub-calls, and 13% against Claude Code) across four diverse long-context tasks while having comparable cost. At a small scale, we post-train the first model around the RLM. Our model, **RLM-Qwen3-8B**, outperforms the underlying Qwen3-8B model by a median of 28% and even approaches the quality of vanilla GPT-5 on three long-context tasks.

User modeling aims to use language models (LMs) to mimic an individual’s behavior from a corpus of past context–action pairs (e.g., conversation turns), enabling the simulation of users in settings like behavioral science, human–AI collaboration, and market research. Recent approaches augment these corpora with synthesized reasoning traces, typically generated by conditioning on both context and action. However, such conditioning constitutes post-hoc rationalization rather than reasoning: the trace is guaranteed to justify the action, but may not encode the underlying latent causal decision paths. We propose Recon, which uses action reconstruction to score reasoning traces by their predictive power: given a context and candidate reasoning, a reconstruction model predicts the action, and reconstruction fidelity determines reasoning quality. Across four domains, Recon achieves a 54.7\% win rate over Backward Synthesis, a standard post-hoc rationalization baseline. Further, we find that training a reasoning synthesis model with rewards derived from Recon improves downstream user modeling performance, achieving a win rate of up to 70.0\% over baselines. We further show that Recon-synthesized reasoning transfers across models, and improves user modeling beyond the reconstruction model. Our work demonstrates that post-hoc rationalization is insufficient for reasoning synthesis, and that useful and interpretable reasoning should naturally elicit the action from the context.

Self-distillation has emerged as a powerful framework for post-training LLMs, where a teacher conditioned on extra information guides a student without it, both from the same model. While this guidance is useful when the student has failed, on successful rollouts, the same mechanism instead overwrites the student's choices and suppresses it's own reasoning. Therefore, we propose reading the original self-distillation signal in reverse: when the student succeeds along a path the teacher would not have predicted, these tokens reflect its self-driven reasoning. Building on this, we propose RLRT (RLVR with Reversed Teacher), which augments GRPO by reinforcing these tokens on correct rollouts. We interpret this as a new form of exploration in RLVR: not uniform diversity, but valuable exploration grounded in the student's own success. Across base, instruction-tuned, and thinking-tuned Qwen3 checkpoints, RLRT substantially outperforms self-distillation and exploration-based baselines, establishing information asymmetry as a new, principled design axis for RLVR.

Image-to-video models often generate videos that remain overly static, compared to text-to-video models. While prior approaches mitigate this issue by weakening or modifying the image-conditioning signal, they often require additional training or sacrifice fidelity to the reference image. In this work, we identify reference-frame dominance as a key mechanism behind motion suppression. We observe that non-reference frames in I2V models allocate excessive self-attention to reference-frame key tokens, causing reference information to be over-propagated across time and suppressing inter-frame dynamics. Based on this finding, we propose DyMoS (Dynamic Motion Slider), a training-free and model-agnostic method that rebalances the attention pathway from generated frames to the reference frame during initial denoising steps. DyMoS leaves both the input image and model weights unchanged and introduces a single scalar parameter for continuous control over motion strength. Experiments across multiple state-of-the-art I2V backbones demonstrate that DyMoS consistently improves motion dynamics while maintaining visual quality and fidelity to the reference image.

Grounding natural language queries to target objects in images requires both high-level reasoning about intentions and precise spatial localization. Multimodal large language models (MLLMs) excel at reasoning about which object is desired given an implicit intention query but turning this understanding into precise bounding box predictions remains challenging. Conversely, specialized detectors such as GroundingDINO offer robust localization capabilities but lack the inherent capacity to understand implicit intentions. Existing hybrid methods couple these components only at the semantic level, such as by passing object names or a single global token to the detector, thereby discarding the rich, dense spatial signals encoded in MLLM attention maps. We propose Reasoning-based Spatial Prior (RSP), a reasoning-based detection framework that converts the MLLM’s internal attention into a learnable spatial prior and injects it directly into the detector. Concretely, we learn input-adaptive gating over attention heads, supervise the fused spatial prior with a SoftIoU localization loss, and integrate it into GroundingDINO via a prior-guided deformable fusion module that biases sampling toward regions highlighted by the reasoning process. Our method consistently outperforms strong MLLM-only, detector-only, and hybrid baselines on the EgoIntention and RIO benchmarks, achieving state-of-the-art performance with substantial gains across both frequent and uncommon categories.

Diffusion-based vision-language-action models (dVLAs) are promising for embodied intelligence but are fundamentally limited in real-time deployment by the high latency of full inference. We propose Realtime-VLA FLASH, a speculative inference framework that eliminates most full inference calls during replanning by introducing a lightweight draft model with parallel verification via the main model's Action Expert and a phase-aware fallback mechanism that reverts to the full inference pipeline when needed. This design enables low-latency, high-frequency replanning without sacrificing reliability. Experiments show that on LIBERO, FLASH largely preserves task performance by replacing many 58.0\,ms full-inference rounds with speculative rounds as fast as 7.8\,ms, lowering task-level average inference latency to 19.1\,ms (3.04$\times$ speedup). We additionally demonstrate effectiveness on real-world conveyor-belt sorting, highlighting its practical impact for latency-critical embodied tasks.

Text-aware image super-resolution (TAISR) aims to recover high-resolution images while preserving textual content. Yet some degraded text cannot be restored from local visual evidence alone: humans often read it by reasoning over image context, logical patterns, and prior knowledge. This raises a fundamental question: does TAISR need reasoning? We answer with two new diagnostic tools: ReasonText, a human-curated benchmark with perception-grounded difficulty levels that separate reasoning-required text from locally recoverable text, and GTTCA, a crop-based recognition accuracy metric that isolates restoration quality from detection errors. Evaluating twelve recent SR models, we find a consistent and substantial performance drop on reasoning-required text, revealing a systematic blind spot in current TAISR designs. To address this limitation, we build on two observations: (1) SR models benefit from ground-truth text labels provided as captions, and (2) modern MLLMs can infer degraded text from LR images through reasoning. These observations motivate Reasoning Transfer via Captioning (RTC), a training-free plug-in that uses an external MLLM as a reasoning module and injects its inferred text into SR models through natural-language captions. Surprisingly, with RTC, real-world SR models can even outperform dedicated text-aware SR models, suggesting that reasoning should be treated as a core design axis for future TAISR.

Large language models (LLMs) show strong performance in multi-document question answering, but their practical deployment is limited by unreliable and unfaithful attribution to supporting evidence. Existing prompting and training-based methods often suffer from hallucinated citations and lack interpretability in how evidence is selected. In this work, we investigate whether attribution signals are inherently encoded within transformer attention mechanisms. We introduce a sensitivity-based diagnostic that identifies a small subset of attention heads, termed **Evidence Heads**, which are highly responsive to perturbations in supporting documents. Through causal interventions and semantic analysis, we show that these heads play a significant role in evidence identification and exhibit alignment with document-level entailment signals. Building on this finding, we propose a training-free Attention-based Attribution framework that extracts evidence signals directly from Evidence Head attention using global and local strategies. Extensive experiments show that our method consistently outperforms strong baselines while remaining lightweight and interpretable. Overall, our results suggest that structured attribution signals are implicitly encoded in LLM attention, and can be effectively leveraged for faithful multi-document reasoning.

Despite recent advances in camera-based online HD map construction, these systems remain susceptible to sensor failures due to their reliance on onboard cameras. In contrast, satellite images remain available independently of real-time sensing conditions as they are cached offline. Existing methods utilizing satellite images, however, primarily leverage them under nominal conditions, without explicitly addressing as redundancy when sensor degrades. To address this gap, we propose ReSMap, a redundancy-aware camera-satellite fusion framework for robust online HD map construction. Our framework explicitly leverages satellite images as a complementary source to improve robustness against camera failures. A Predict-then-Fuse (PTF) module fuses camera and satellite BEV features by using each branch's own prediction confidence as a per-pixel reliability weight, with no learned router. An Object-centric cooperative Multi-modal Query (OCMQ) decoder then propagates this isolation to the query level: each modality cross-attends only to its own BEV, and per-modality updates are merged through a learned per-token gate. Our model achieves state-of-the-art performance on nuScenes across all splits and range settings. More importantly, it also sets the new state-of-the-art under camera-failure scenarios, demonstrating that satellite images serve not only as a strong prior but as practical redundancy for robust online HD mapping.

Vision-Language-Action (VLA) models remain brittle in long-horizon, contact-rich manipulation because success-only imitation provides little supervision for execution drift, while failed rollouts are often discarded. We introduce RePO-VLA, a recovery-driven policy optimization framework that assigns distinct roles to success, recovery, and failure trajectories. RePO-VLA first applies Recovery-Aware Initialization (RAI), slicing recovery segments and resetting history so corrective actions depend on the current adverse state rather than the preceding failure. It then learns a Progress-Aware Semantic Value Function (PAS-VF), aligning spatiotemporal trajectory features with instructions and successful references. The resulting labels salvage useful failure prefixes via reliability decay, while low-value labels mark drift and terminal breakdowns, teaching differences among nominal, failed, and corrective actions. The data engine turns adverse states into planner-generated or human-collected corrective rollouts, teaching recovery to the success manifold. Value-Conditioned Refinement (VCR) trains the policy to prefer high-progress actions. At deployment, a fixed high value ($v=1.0$) biases actions toward the learned success manifold without online failure detectors or heuristic retries. We introduce FRBench, with standardized error injection and recovery-focused evaluation. Across simulated and real-world bimanual tasks, RePO-VLA improves robustness, raising adversarial success from 20\% to 75\% on average and up to 80\% in scaled real-world trials.

Video world models aim to simulate environmental dynamics under interaction. However, most existing approaches entangle transition reasoning with pixel-level generation within a single model, thereby limiting narrative continuity and long-horizon coherence. We reformulate video world modeling as an agentic process that synergizes reasoning and generation, and present ReGen, which couples a streaming vision-language planner with a causal diffusion generator. At each step, the planner infers a structured action-state specification from history, while the generator realizes it as the next video segment. The planner further verifies the outcome and decides to continue, end, or re-generate, forming a closed-loop sense–plan–act–verify pipeline. To support this formulation, we carefully construct a large-scale action-grounded video dataset with over 2M segments that instantiates the planner–generator interface, and introduce action-grounded reward alignment that turns the action specification into an enforced reward and mitigates cross-segment drift. Experiments show that ReGen improves long-horizon narrative coherence over prior work, supporting autonomous long-horizon rollouts without step-by-step user intervention.