NeurIPS 2026 Papers

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.

In the field of Neural Combinatorial Optimization (NCO), Reinforcement Learning (RL) stands out for its ability to naturally enforce complex constraints via its auto-regressive decision pattern. To mitigate the longstanding challenges, e.g., reward sparsity and sample inefficiency in the vast combinatorial action space which causes ineffective training, subpar performance, and poor scalability, we propose CORectifier, a novel NCO solver learned under the hierarchical gated rectification mechanism to regularize the arbitrary exploration: partial policy-predicted actions in a trajectory are probabilistically replaced with high-quality segments from reference solutions. CORectifier prompts the model with tri-level optimal signals, operating at the batch, instance, and sub-instance levels with fragments of diverse lengths injected at random decision steps. This Rectified RL (RRL) paradigm helps develop optimality-aware and sample-efficient RL learners while maintaining their sequential-decision manner for constraint satisfaction, delivering a new perspective to hybridize RL and SL/IL for NCO with improved utility rate of limited supervision. Sufficiently extensive experiments on Traveling Salesman Problem (TSP), Asymmetric TSP (ATSP), Prize-Collecting TSP (PCTSP), Capacitated Vehicle Routing Problem (CVRP), Knapsack Problem (KP), Job-Shop Scheduling Problem (JSSP), and Single Machine Total Weighted Tardiness Problem (SMTWTP), across synthetic and real-world benchmarks, show superior quality than RL baselines by up to 59.7%.

Evaluating video generation models is notoriously challenging; existing metrics either fail to align with human judgment or rely on costly, unscalable human ratings. In this work, we introduce \textbf{Video Preference Score} (VPS), a fully automated video preference score designed to directly model human judgment and accurately recreate large-scale, Elo-based video arenas. Trained on a newly collected dataset that augments generated videos with retrieved real-world footage, VPS evaluates paired videos against complex text prompts to predict human preference. Experiments demonstrate that VPS correlates highly with human consensus and generalizes zero-shot across unseen models and benchmarks. Furthermore, we showcase the utility beyond evaluation: as a plug-in reward model, VPS seamlessly integrates with Flow-GRPO to improve WAN-2.1's generative quality by 51 Elo points. Overall, VPS provides a scalable and generalizable solution for both evaluating and aligning modern video generation models.

We study the sample complexity of noisy one-bit compressed sensing for signals drawn from a prior distribution. By characterizing the effective distributional complexity of the prior via its approximate covering number, we prove that posterior sampling achieves accurate recovery with high probability when the number of measurements scales with the logarithm of the approximate covering number, up to a one-bit separation gap factor. This upper bound is robust to learned prior mismatch. Specifically, we show that posterior sampling with an approximate prior remains reliable, provided that the learned prior distribution is sufficiently close to the true signal distribution in Wasserstein distance. In addition, we establish a sample complexity lower bound for any reliable method of noisy one-bit compressed sensing, showing that our upper bound is nearly matched in its main prior dependent term. To approximate the ideal posterior sampling process for real world scenarios, we instantiate posterior sampling through a plug-and-play algorithm with diffusion priors. Experiments on the FFHQ and ImageNet datasets demonstrate the effectiveness of our proposed approach.

Hierarchical clustering seeks to uncover nested structure in data by constructing a tree of clusters, whose deeper levels reveal increasingly fine-grained relationships. However, traditional hierarchical clustering methods always return a hierarchy, even when the data contain no meaningful hierarchical structure. Moreover, agglomerative linkage algorithms are highly sensitive to the choice of linkage function. In this paper, we revisit a classical notion of well-separated clusters, which we call valid clusters. The collection of all valid clusters forms a hierarchy, known as the Apresjan hierarchy. This hierarchy is the finest hierarchy composed only of valid clusters: it need not be binary, and it collapses to a star tree when no nontrivial valid clusters exist. Our main contribution is to characterize when the Apresjan hierarchy can be recovered from linkage-based algorithms. We propose a two-step procedure that first constructs a binary hierarchy using an agglomerative linkage method and then prunes all clusters that violate the validity condition. We give necessary and sufficient conditions on the linkage function under which this procedure exactly recovers the Apresjan hierarchy. Consequently, all linkage rules satisfying these conditions yield the same pruned hierarchy. In particular, single, complete, and average linkage satisfy the conditions, whereas Ward's linkage does not.

Personalization systems encode long-horizon, evolving user trajectories into compressed preference states that downstream task heads consume for prediction or generation. This compression is difficult because preference trajectories contain stable long-term interests, transient short-term shifts, and episodic bursty interests. We argue that a practical route is not to design another task-specific encoder or incur costly host-encoder finetuning, but to correct the under-expressed compressed state before the downstream task head consumes it. We propose $\texttt{REPAIR}$ as a selective *pre-head state-repair method* for frozen personalization encoders. It plugs into frozen hosts, uses event representations from the forward pass, and avoids a second pass over the raw history. $\texttt{REPAIR}$ compares the compressed state with per-timestep event representations, estimates state-relative corrective evidence, resolves it through long-term, short-term, and episodic temporal regimes, and retains only the most corrective signals to form a compact correction. We evaluate $\texttt{REPAIR}$ across recommendation and generation personalization settings using MovieLens, MIND, PENS, and Amazon Reviews 2023. Across recommendation tasks, $\texttt{REPAIR}$ improves the evaluated frozen hosts. On MovieLens, it improves the highest-scoring frozen host in our instantiated baseline suite by +1.27/+1.54 MRR/nDCG@5; on MIND, it improves the highest-scoring frozen host by +4.41 MRR; and on Amazon product recommendation, it improves two SOTA baselines by +15.33 and +17.91 MRR. It also outperforms budget-matched post-head correctors, and the full long-, short-, and episodic repair remains the strongest temporal variant. On PENS personalized headline generation, $\texttt{REPAIR}$ improves subjectivity-sensitive generation most when the decoder is preference-coupled, with average PerSEval gains of +14.62% and +22.19% for two representative preference-coupled decoders. Relative to frozen predictive encoders, deployment overhead is 31--34% memory and 38--44% latency.

When benchmark examples overlap with a model’s training data, benchmark contamination can substantially inflate reported evaluation metrics, yet training provenance is often unavailable. We study post-hoc recovery of uncontaminated evaluation metrics in supervised tabular classification and regression using only prediction–label pairs from a potentially contaminated benchmark. We introduce MILAN (Metric Integrity via Leakage-Aware Normalization), a meta-trained framework that models a contaminated benchmark as a mixture of training-origin and clean evaluation examples and estimates soft posterior weights for the clean component to correct evaluation metrics. MILAN supports a broad class of metrics through weighted estimators and metric-specific correction procedures, and can optionally incorporate contamination-rate priors and sample-wise difficulty scores. To quantify uncertainty, MILAN additionally provides split-conformal prediction intervals with marginal coverage across exchangeable recovery problems. Across a large suite of simulated recovery problems with held-out datasets and model families, MILAN consistently improves clean-metric recovery over thresholding, clustering, mixture-model, anomaly-detection, and membership-inference baselines. In a polygenic risk score case study using UK Biobank evaluation data, MILAN improves agreement with independent FinnGen reference metrics. We provide a scikit-learn-style implementation at \url{hidden-for-submission}.

Robotic manipulation policies have achieved remarkable success in controlled settings, but they catastrophically fail when execution errors occur---dropped objects, collisions, misaligned grasps. Despite the critical importance of error recovery for real-world deployment, no existing benchmark systematically evaluates manipulation policies' ability to recover from such errors. We introduce RecoverBench, the first systematic benchmark for error recovery in robotic manipulation. We propose a taxonomy of 12 Error Skills organized into 5 Recovery Behavior Groups (RBGs), spanning 24 error subtypes across 2 difficulty degrees. Our Error Skill pipeline automatically generates 1,360 reproducible error scenes from clean demonstration trajectories across 6 manipulation tasks. Each error scene includes complete simulation state snapshots, environment fingerprints, and RNG states for deterministic reproduction. Beyond evaluation, RecoverBench also supports recovery-demo collection and augmentation, and we show that this recovery supervision improves recovery performance. We evaluate 4 representative manipulation policies on RecoverBench. Even the best-performing policy reaches only 48.6% average recovery success, ranging from 32.0% on threading to 76.7% on stack, showing that current methods still lack robust recovery. These results highlight the need for dedicated recovery evaluation and position RecoverBench as a benchmark for improving recovery robustness in manipulation policies. Code, data, and evaluation protocols will be released.

World model-based policy evaluation is a practical proxy for testing real-world robot control by rolling out candidate actions in action-conditioned video diffusion models. As these models increasingly adopt latent diffusion modeling (LDM), choosing the *right latent space* becomes critical. While the status quo uses autoencoding latent spaces like VAEs that are primarily trained for pixel *reconstruction*, recent work suggests benefits from pretrained encoders with representation-aligned *semantic* latent spaces. We systematically evaluate these latent spaces for action-conditioned LDM by comparing six reconstruction and semantic encoders to train world model variants under a fixed protocol on BridgeV2 dataset, and show effective world model training in high-dimensional representation spaces with and without dimension compression. We then propose three axes to assess robotic world model performance: visual fidelity, planning and downstream policy performance, and latent representation quality. Our results show visual fidelity alone is insufficient for world model selection. While reconstruction encoders like VAE and Cosmos achieve strong pixel-level scores, semantic encoders such as V-JEPA 2.1 (strongest overall on policy), Web-DINO, and SigLIP 2 generally excel across the other two axes at all model scales. Our study advocates semantic latent space as stronger foundation for policy-relevant robotics diffusion world models.

Masked diffusion language models (MDLMs) generate text by unmasking tokens in parallel and have recently emerged as alternatives to autoregressive language models. They can be viewed as parallel decoders trained with a position-wise cross-entropy (CE) loss, the same setup as non-autoregressive translation (NAT). In NAT, CE-trained parallel decoders have been argued to be sensitive to small positional shifts, since CE penalizes them harshly. We ask whether CE-trained MDLMs are similarly sensitive to such shifts under iterative decoding. To probe this, we apply a controlled intervention that introduces them during decoding. On LLaDA-8B-Instruct with Arena-Hard, displacing as little as 1\% of generated tokens by one position substantially reduces win rates against the unintervened model, showing that MDLMs are sensitive to such small shifts under iterative parallel decoding. Motivated by this, we adapt connectionist temporal classification (CTC), an alignment-flexible objective known to mitigate it there, to MDLM supervised fine-tuning. By relaxing the strict position-wise match that CE imposes, CTC gives the loss room to absorb small positional shifts; concretely, we modified CTC objective to use a special \ token that absorbs positional uncertainty between target tokens and output positions, and a updated collapse map that preserves target surface forms. Across four open-ended generation benchmarks, the resulting model consistently improves over both the original model and a matched cross-entropy-trained baseline, with statistically significant gains on all four. These results identify training-side alignment flexibility as a useful design dimension for MDLM SFT, complementary to the inference-time approaches explored in prior work.

General-purpose behavior models increasingly internalize broad procedural knowledge at training time, making adaptation less a problem of learning entirely new behaviors than of reconfiguring existing competence for new goals, object relations, and skill interfaces. We study this as reconfigurability, the ability to realize new behaviors by restructuring components and relations cost-efficiently, under *sparse governing information*, where a task specification provides high-level goals or abstract directives but omits the concrete bindings and compositions required for execution. We propose RecPo, a framework that represents skills task-agnostically through multiple procedure views and materializes a new task by composing only the transition relations consistent with the task-relevant views. RecPo maintains a procedure knowledge base of local transition relations within each view. At deployment time, it interprets a task demonstration through the same views, converts observed view-level changes into a compositional query, and resolves the query into an executable skill sequence. On an extended Franka Kitchen benchmark with held-out long-horizon tasks outside the stored compositions, RecPo achieves 91.7\% ordered subtask completion, substantially outperforming non-reconfigurable skill retrieval baselines. Real-robot case studies further show deployment-time adaptation through recomposition rather than task-specific retraining.

Ensuring high returns while providing strong safety guarantees in offline imitation learning (IL) is a fundamental challenge: existing approaches either rely on explicitly specified cost functions, which are difficult to design for complex real-world safety constraints, or require expert demonstrations that are both safe and high-performing, which are often too costly to collect. We introduce *dual-expert imitation learning*, which leverages two complementary data sources: (i) safety-compliant demonstrations that faithfully follow safety guidelines but may be suboptimal, and (ii) performance-oriented demonstrations that achieve high returns but may violate safety guidelines. To address this problem, we propose DexDICE (**D**ual-**ex**pert offline imitation learning via stationary **DI**stribution **C**orrection **E**stimation). Instead of conventional distribution constraints, DexDICE combines DICE-style offline learning with safe support constraints, enabling agents to exploit high-return behaviors while remaining within safety-compliant regions. We evaluate DexDICE on a re-curated DSRL benchmark and validate it on a real-world mobile-robot navigation task trained directly from human teleoperation demonstrations, where it significantly outperforms offline IL baselines and remains robust to data scarcity and quality degradation.

Deep decarbonization is transforming power grids into safety-critical, low-carbon cyber-physical systems, where AI controllers must jointly manage economic efficiency, carbon reduction, and operational security. This operational energy trilemma is challenging for reinforcement learning: cost and carbon objectives are naturally optimized in expectation, whereas security violations are rare, heavy-tailed events whose consequences can be catastrophic and therefore require explicit tail-risk control. Existing safe reinforcement learning methods either constrain safety only in expectation or rely on light-tailed approximations, which can underestimate rare but severe grid violations. We propose a heterogeneous risk-constrained reinforcement learning framework for low-carbon grid operation. The key idea is to assign different risk semantics to different objectives: safety-critical security constraints are enforced through Conditional Value-at-Risk (CVaR), while economic and carbon-related objectives are optimized in expectation to preserve operational flexibility. To better capture extreme scenarios, a mixture distribution model is introduced to characterize heavy-tailed constraint violations. We further develop a residual policy learning scheme built on a distributionally robust chance-constrained reference module: the robust module provides a feasible and economically efficient baseline policy, and the learned residual refines it toward lower-carbon operation while respecting tail-risk bounds. Experiments on power-system operation tasks show that the proposed framework reduces economic cost by over 25% compared with expectation-based CMDP baselines and improves safety by two orders of magnitude over Gaussian-tail methods. These results suggest that risk-aware reinforcement learning can support reliable decarbonization by reconciling efficiency, sustainability, and security within a unified decision-making framework.

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.

Sequential recommenders need long histories, but full-prefix attention scales with user lifetime. We argue that an effective memory mechanism should provide space invariance, semantic fidelity, and lifelong evolvability, which existing fixed-window or capped-memory methods do not jointly satisfy. We propose RecMem, a recurrent compression framework that folds history segments into M forget-gated memory slots. RecMem has a history-length-independent rollout error bound, and its fixed-capacity memory is trained by next-item prediction to preserve useful signals. On MerRec, RecMem uses 5.6x fewer decoder-side tokens than full attention while retaining 88–97% of Recall@10–200. Compared with a compressed truncation baseline at similar decoder cost, RecMem improves Recall@50+ by 4–6 percentage points by compressing the full history rather than discarding old interactions. It also remains stable over 1,800 test steps.

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.

Large Reasoning Models (LRMs) leverage explicit reasoning traces, known as Chain-of-Thought (CoT) reasoning, to decompose complex problems into intermediate steps before deriving final answers. However, these reasoning traces introduce unique safety challenges: harmful content can be embedded in intermediate steps even when final answers appear benign. Our study reveals that existing state-of-the-art moderation tools experience significant performance degradation on CoT moderation, with F1 scores dropping by up to 35.2% compared to traditional answer moderation. To address these challenges, we present ReasoningShield, a comprehensive benchmark and a suite of strong lightweight models for CoT safety moderation. Our benchmark includes 9.2K annotated Query-CoT pairs with structured stepwise risk analysis, consisting of a 7K-sample training set and a 2.2K human-annotated test set, spanning 10 risk categories, 3 safety levels, and 8 LRMs with diverse reasoning paradigms. To further demonstrate the utility of the benchmark, we develop lightweight models with 1B and 3B parameters using a two-stage training strategy. Our models achieve 91.8% F1 on our benchmark, substantially outperforming leading tools such as LlamaGuard-4 by 35.6% and commercial models such as GPT-4o by 15.8%, while generalizing effectively across diverse reasoning paradigms and unseen scenarios. All resources, including the dataset, code, and models, are released at https://anonymous.4open.science/r/ReasoningShield.

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.