NeurIPS 2026 Papers

We study transfer learning for a linear regression task using several least-squares pretrained models that can be overparameterized. We formulate the target learning task as optimization that minimizes squared errors on the target dataset with penalty on the distance of the learned model from the pretrained models. We analytically formulate the test error of the learned target model and provide the corresponding empirical evaluations. Our results elucidate when using more pretrained models can improve transfer learning. Specifically, if the pretrained models are overparameterized, using sufficiently many of them is important for beneficial transfer learning. However, the learning may be compromised by overparameterization bias of pretrained models, i.e., the minimum L2-norm solution's restriction to a small subspace spanned by the training examples in the high-dimensional parameter space. We propose a simple debiasing via multiplicative correction factor that can reduce the overparameterization bias and leverage more pretrained models to learn a target predictor.

Diffusion policies have emerged as a powerful policy class for offline reinforcement learning (RL) due to their ability to model complex, multimodal action distributions through iterative denoising. However, in offline RL, expressivity alone is not enough. Generated actions are supposed to remain within value-reliable regions of the action space to avoid extrapolation error. We find that standard diffusion policies often produce unstable denoising trajectories whose intermediate and final action candidates drift toward weakly supported regions, making critic-guided action selection unreliable and degrading teacher quality for one-step distillation. This issue is particularly pronounced on mixed-quality offline datasets, where suboptimal behaviors amplify trajectory instability. To address this problem, we propose Trajectory Consistency Rectification (TCR), a framework designed to promote denoising stability during both training and inference. At inference time, TCR aggregates locally consistent, high-value action candidates across denoising steps to suppress critic-unreliable outliers. During training, it adaptively reweights denoising steps to improve trajectory-level coherence. We show that this stabilization provides a more reliable teacher for one-step flow-matching distillation. Extensive experiments on D4RL and OGBench show that TCR consistently improves over diffusion-policy baselines, with significant gains in sparse-reward domains like AntMaze and Adroit. Moreover, one-step policies distilled from TCR achieve competitive performance among efficient single-step methods. These results suggest that enhancing the temporal consistency of denoising trajectories is a key ingredient for robust offline diffusion reinforcement learning.

Transformers perform inference by iteratively transforming token representations across layers. This layerwise computation has been studied empirically, and recent mean-field theories of Transformer dynamics explain how attention can drive token distributions toward clustering. However, existing mean-field analyses largely treat model parameters as prescribed, leaving open how training reshapes this clustering picture. We study this question in a noisy mean-field Transformer in which only a parameter-linear FFN is trained under $L^2$ regularization. We find and analyze a training-induced phase in the dynamics: after initially following attention-driven clustering, the token distribution can leave the clustered regime near the final layers. Our mathematical analysis is based on an entropy-regularized interaction energy that captures the clustering bias of attention. More broadly, our results point toward a training-aware mean-field theory of Transformer dynamics, in which training and inference dynamics are treated together.

Chain-of-thought (CoT) monitoring is one of the most promising tools we have for detecting model misbehavior, but its effectiveness depends on models faithfully externalizing their reasoning. Motivated by this vulnerability, we study whether monitor-aware models are capable of obfuscating their reasoning to evade detection. We use synthetic document finetuning to expose eight models to realistic pre-training-style documents describing a CoT monitor and find that monitor-aware models consistently achieve higher rates of undetected misbehavior compared to unaware controls. This effect is weaker but still present on a harder agentic task. We also show that CoT controllability, a model's ability to reshape its own reasoning trace under an imposed constraint, is closely correlated with obfuscation success across the eight models studied ($r=0.800$, $p=0.017$). Monitor-aware models placed under equal reinforcement learning optimization pressure also learn to reward-hack without triggering a CoT monitor substantially faster than unaware controls. Together, these results suggest that knowledge of monitoring combined with high CoT controllability poses a risk to CoT-based monitoring.

Many emerging agentic paradigms require agents to collaborate with one another (or people) to achieve shared goals. However, existing approaches to learning policies for such collaborative problems produce brittle solutions that fail when paired with new partners. We attribute these failures to a combination of free-riding during training and a lack of strategic robustness. To address these problems, we study the concept of strategic risk aversion and interpret it as a principled inductive bias for generalizable cooperation with unseen partners. While strategically risk-averse players are robust to deviations in their partner's behavior by design, we show that, in collaborative games, they also (1) can have better equilibrium outcomes than those at classical game-theoretic concepts like Nash, and (2) exhibit less or no free-riding. Inspired by these insights, we develop a multi-agent reinforcement learning (MARL) algorithm that integrates strategic risk aversion into standard policy optimization methods. Our empirical results across collaborative benchmarks (including an LLM collaboration task) validate our theory and demonstrate that our approach consistently achieves reliable collaboration with heterogeneous and previously unseen partners across collaborative tasks.

Inference-time scaling provides a promising path for improving agentic reasoning performance, and one widely adopted form is sample-then-select: first sample a pool of candidate reasoning trajectories, then select from the resulting candidates. While this keeps scaling lightweight without updating the policy, it exposes a training-deployment mismatch. Standard trajectory-wise Reinforcement Learning (RL) objectives optimize normalized signals for individual trajectories, whereas choosing from a large candidate pool with the policy's own selection signal is inherently set-conditional: it requires calibrated ranking and confidence aggregation across correlated trajectories. To address this challenge, we propose AutoPortfolio, a set-conditional policy alignment framework for efficient and effective sample-then-select scaling. AutoPortfolio builds a tractable environment-informed target on the sampled candidates, and uses its induced Plackett-Luce (PL) ranking to calibrate the policy which trajectories to prioritize within that set. It trains the policy with a balanced PL-motivated objective that reallocates probability mass among competing trajectories, preserving reward-aligned alternatives while sharpening away from low-quality distractors. Aligned with our policy training, our lightweight inference-time Mass Aggregation scaling adaptively perceives confidence from candidate trajectories and selects the most promising answer(s). Theoretically, we show that reducing our set-conditional loss narrows the discrepancy between policy-induced and environment-informed PL rankings on the sampled set. Empirically, across 12 challenging reasoning benchmarks, AutoPortfolio improves over strong agentic RL baselines, with the clearest gains in larger candidate pools where the set-level selection is crucial.

Recent progress in computer vision has produced a wide range of powerful specialized models for detection, segmentation, counting, and other visual tasks. However, these models are usually optimized for isolated task formulations, making it difficult to directly support general-purpose visual intelligence, especially when a task requires complex language understanding and dense small-object perception. In this paper, we propose VisHarness, a trainable visual agent that decouples high-level perception, reasoning, and decision-making from low-level task execution. Instead of training a model to solve a specific visual task, VisHarness learns to harness a set of carefully designed heterogeneous visual experts. This paradigm preserves the general intelligence of the agent while fully leveraging the precision advantages of specialized visual models in concrete visual tasks. With only lightweight training, VisHarness learns a generalizable visual expert-harnessing policy and can solve common fundamental vision tasks under various complex conditions through multi-turn interactions with visual expert models. To enable efficient on-policy reinforcement learning training in a live environment, we introduce dynamic visual memory archiving, which mitigates the rapidly accumulating visual-token overhead caused by multi-turn interactions with visual expert models. Experiments on four representative benchmarks covering reasoning segmentation, generalized referring segmentation, dense small-object detection, and referring counting demonstrate that VisHarness substantially outperforms existing general-purpose models and achieves competitive or superior performance compared with task-specific models.

Reinforcement learning (RL) has become essential for post-training large language models (LLMs) in reasoning tasks. While scaling rollouts can stabilize training and enhance performance, it introduces substantial computational overhead. In algorithms like GRPO, multiple rollouts per prompt incur prohibitive costs, as a large portion of prompts provide negligible gradients and are thus of low utility. This raises a key question: \textit{how to identify high-utility prompts before an expensive rollout?} Our experimental analysis reveals that sample utility is non-uniform and dynamic: the strongest learning signals concentrate at the ``learning edge'', the intersection of intermediate difficulty and high response entropy, which shifts throughout training. Motivated by this observation, we propose HIVE, a history-informed and online-verified prompt selection framework for data-efficient RL training. HIVE first uses historical reward statistics and response entropy as a cheap prior to filter candidate prompts, and then employs prompt entropy as a real-time proxy to prune instances with stale utility. Across multiple reasoning benchmarks and base models, HIVE maintains reasoning accuracy with substantial rollout cost reduction, achieving up to 2.3× total training speedup.

Optimal pre-training data mixtures are vital for Large Language Models (LLMs), yet post-RL reasoning performance varies significantly across base models. This suggests that identifying "RL-friendly" pre-training data mixtures is a critical but under-explored prerequisite. This motivates us to ask: Which specific data mixtures in pre-training enhance post-RL effectiveness, and how can we evaluate a base model's suitability for RL fine-tuning? Through large-scale experiments with proxy models, we observe that (1) domain-specific data correlation strongly predicts post-RL performance, and (2) the average perplexity of the base model on $k$ RL-generated responses, $\text{avg}(k\text{-ppl})$, is a more stable and correlated predictor than traditional metrics. Building on these insights, we propose BridgeRL, a framework that treats finding optimal mixing ratios as a regression task, using $\text{avg}(k\text{-ppl})$ as the fitting objective. We train 1M/60M-parameter proxy models for regression and scale the findings to 1B/3B-parameter models. Results across online and offline RL algorithms demonstrate that BridgeRL-optimized mixtures consistently yield superior base models for RL fine-tuning.

Multimodal deep search requires an agent to solve open-world problems by chaining search, tool use, and visual reasoning over evolving textual and visual context. Two bottlenecks limit current systems. First, existing tool-use harnesses treat images returned by search, browsing, or transformation as transient outputs, so intermediate visual evidence cannot be re-consumed by later tools. Second, training data is usually built by fixed curation recipes that cannot track the target agent's evolving capability. To address these challenges, we first introduce a visual-native agent harness centered on an image bank reference protocol, which registers every tool-returned image as an addressable reference and makes intermediate visual evidence reusable by later tools. On top of this harness, On-policy Data Evolution (ODE) runs a closed-loop data generator that refines itself across rounds from rollouts of the policy being trained. This per-round refinement makes each round's data target what the current policy still needs to learn. The same framework supports both diverse supervised fine-tuning data and policy-aware reinforcement learning data curation, covering the full training lifecycle of the target agent. Across 8 multimodal deep search benchmarks, ODE improves the Qwen3-VL-8B agent from 24.9% to 39.0% on average, surpassing Gemini-2.5 Pro in standard agent-workflow setting (37.9%). At 30B, ODE raises the average score from 30.6% to 41.5%. Further analyses validate the effectiveness of image-bank reuse, especially on complex tasks requiring iterative visual refinement, while rollout-feedback evolution yields more grounded SFT traces and better policy-matched RL tasks than static synthesis.

We present the first theoretical guarantees for differentially private online reinforcement learning (RL) with general function approximation, extending beyond prior work restricted to tabular and linear settings. Our approach combines a batched policy update scheme with the exponential mechanism, together with a novel regret analysis. We show that, even under general function approximation, the regret in the model-free setting under differential privacy matches the state of the art for the linear case, scaling as $\widetilde{O}(K^{3/5})$, where $K$ denotes the number of episodes. As an important by-product, we also establish the first regret bound for online RL with batch update that depends on the standard complexity measure of **coverability**, complementing existing results based on a newly introduced Eluder-Condition class. In addition, we uncover fundamental gaps in recent results for private RL with linear function approximation, thereby clarifying its landscape.

Proximal Policy Optimization (PPO) and it's variants are the most widely used policy-gradient methods in reinforcement learning, yet the role of its actor update mechanism is still poorly understood theoretically. In particular, standard PPO combines clipped surrogate gradients, multiple epochs of minibatch updates, and reuse of rollout data, but existing convergence analyses do not fully capture this update structure. In this work, we theoretically study PPO policy updates with symmetric clipping from a policy-gradient perspective and interpret its actor updates as a cyclic biased-gradient method with sample reuse and random reshuffling. Our first contribution is a clean formalization of PPO clipped actor updates through surrogate gradients that approximate the true policy gradient. Using the performance difference lemma, we prove a linear bias bound which quantifies how the surrogate gradient drifts as the policy moves away from the sampling policy. Our second contribution is a convergence analysis of cyclic surrogate-gradient ascent, showing that additional PPO-style biased updates can improve progress under conservative learning rates without requiring additional samples. Finally, we analyze the stochastic minibatch version with reshuffling and obtain convergence-to-stationarity guarantees under standard smoothness assumptions and a bounded critic-bias condition. Overall, our results provide a theoretical interpretation of PPO’s multi-epoch actor updates: extra clipped surrogate steps introduce bias, but can still improve optimization efficiency by compensating for small, stable step sizes through sample reuse.

Online robust zero-sum Markov games pose a coupled operator-learning and game-solving problem: from nominal interaction data, the learner must recover a worst-case minimax Bellman operator while maintaining strategic coverage and controlling stage-game approximation error. Existing robust reinforcement-learning theory does not yet provide such an operator-learning perspective beyond specialized multi-agent settings. We develop an optimistic dual robust Bellman framework that combines dual robust Bellman fitting, optimistic minimax planning, backward-validated confidence sets, and exploiter-mix data collection. Our analysis is organized around the robust Bellman residuals induced by the candidate value class and yields a regret reduction that separates exploration complexity, robust operator-estimation error, stage-game solution error, and misspecification. Under explicit interpretable assumptions, we obtain conditional theorem interfaces for general and projected function approximation, together with closed results in several structured regimes, including tabular, epochwise/cross-fit linear, self-normalized pure-online linear / finite-rank kernel, and projected RKHS truncation. We also provide numerical experiments that validate the theoretical predictions.

Generating codes from UI screenshots has recently benefited from MLLMs, yet existing methods centered on HTML targets and transfer poorly to multi-frameworks such as React, Vue, and Angular. However, unlike HTML, which does not entail compilation issues, generating executable code in multi-framework settings is more challenging due to framework-specific differences in syntax and failure modes, as well as the need to maintain global consistency across files. To address this problem, we study executable multi-framework front-end code generation and propose MESCoder, a tree-structured self-corrective generation framework. Our method first extracts a structural scaffold from the screenshot and then constructs sub-trees and assembly relations step by step, transforming large project generation into a sequence of localized decisions. On top of this scaffold, we introduce a unified Project Agent that continuously revises its own partial project under runtime sandbox feedback and dynamically chooses between local repair and cluster level repair. For training, we adopt a two-stage strategy that first initializes the policy offline to learn basic multi-framework generation and repair priors, and then refines it with online reinforcement learning to optimize long horizon executability and final page quality. The resulting framework models multi-framework executability as a conditioned project state transition process and improves the stability and quality of complex front end generation through self correction. Experiments show that our method improves executability, while narrowing the visual-fidelity gap to HTML-based generation.

Building embodied agents that can perceive, reason, and act across diverse 3D worlds is a central research direction toward artificial general intelligence. Yet most existing work on embodied agents remains confined to specific tasks in a single scene and transfers poorly to new environments, leaving general embodied agents for open-world settings an important but under-explored direction. We refer to such agents as Embodied World Agents (EWAs), whose defining property is the ability to understand diverse visual environments and execute general actions in open-world conditions. To advance research on EWAs, we first release EmbodiedWorld-200K, a large-scale dataset of about over 200K samples for open-world embodied planning that spans diverse scenes. Each example is equipped with multi-level annotations, ranging from low-level camera-motion trajectories to high-level semantic instructions, and is accompanied by unified evaluation splits and task-specific metrics, providing a shared substrate for offline pretraining and fair evaluation of EWAs. Beyond offline supervision, a defining characteristic of an EWA is its interaction with the environment; training solely on offline data is therefore insufficient, especially for generalizing to scenes not covered in the training distribution. Building dedicated real interactive environments, however, is often prohibitively expensive. To address this, we propose a novel training scheme that employs a video world model as a low-cost environment proxy and performs interactive reinforcement-learning (RL) post-training of the EWA within it. Empirically, pretraining on EmbodiedWorld-200K substantially improves the open-world embodied planning ability of the EWA, and further RL post-training inside a world model yields clear gains on out-of-distribution scenes. The full dataset, baseline checkpoints, and the accompanying training and evaluation code for supporting future research on EWAs are publicly available at https://xiaokunfeng.github.io/EmbodiedWorld-200K/.

Multimodal Large Language Models (MLLMs) have achieved remarkable success in open-vocabulary perceptual tasks, yet their ability to solve complex cognitive problems remains limited, especially when visual details are abstract and require visual memory. Current approaches primarily scale Chain-of-Thought (CoT) reasoning in the text space, even when language alone is insufficient for clear and structured reasoning, and largely neglect visual reasoning mechanisms analogous to the human visuospatial sketchpad and visual imagery. To mitigate this deficiency, we introduce Cognitive Supersensing, a novel training paradigm that endows MLLMs with human-like visual imagery capabilities by integrating a Latent Visual Imagery Prediction (LVIP) head that jointly learns sequences of visual cognitive latent embeddings and aligns them with the answer, thereby forming vision-based internal reasoning chains. We further introduce a reinforcement learning stage that optimizes text reasoning paths based on this grounded visual latent. To evaluate the cognitive capabilities of MLLMs, we present CogSense-Bench, a comprehensive visual question answering (VQA) benchmark assessing five cognitive dimensions. Extensive experiments demonstrate that MLLMs trained with Cognitive Supersensing significantly outperform state-of-the-art baselines on CogSense-Bench and exhibit superior generalization on out-of-domain mathematics and science VQA benchmarks, suggesting that internal visual imagery is potentially key to bridging the gap between perceptual recognition and cognitive understanding. We will open-source the CogSense-Bench and our model weights.

We propose a contextual cascading bandit model in which the click probability at each position depends on both the displayed item and the topics of previously shown items. This captures how previously displayed topics affect later click probabilities, allowing different topic orders to induce different user responses. Unlike standard cascading bandits, the optimal cascade is no longer determined only by the selected set of items, since the ordering itself affects the expected reward. To address the resulting planning challenge, we reformulate cascade construction as a Markov decision process whose state summarizes the previously selected topics. Based on this formulation, we develop an optimism-based algorithm and prove a regret upper bound of $\widetilde{\mathcal O}(\bar p^{\frac{K-1}{2}}d\sqrt{T})$, where $d$ is the number of unknown parameters, $T$ is the horizon, $K$ is the cascade length, and $\bar p<1$ upper bounds the no-click probability of an examined item. A key implication is that the regret decreases with the cascade length $K$, a phenomenon that had not been established even in order-insensitive contextual cascading bandits. We further provide a local regret lower bound showing that this decreasing dependence on $K$ is intrinsic, and validate our theory through experiments with topic-order-dependent click probabilities.

Large language models (LLMs) excel at natural language processing but struggle to interact with external environments. Tool learning provides a promising way to extend LLMs into actionable agents, where tool selection is a critical prerequisite for successful tool use. Existing work often assumes a small or predefined set of tools, leaving large-scale tool selection underexplored. Real-world repositories contain a vast and diverse array of tools, making it difficult for LLMs to effectively search, distinguish, and compose tools under context-length constraints. We identify large-scale tool selection as a new challenge for agentic reinforcement learning, highlighting that existing RL methods for knowledge-based question answering are inadequate for selecting tools while considering compatibility. To address this challenge, we propose ToolSearcher, a novel RL framework for effective multi-turn search and fine-grained optimization in large-scale tool selection. Specifically, we introduce category-constrained tool discrimination to improve the model's ability to distinguish functionally similar tools, event-level search modeling to explicitly optimize the discovery of target tools during multi-turn search, and trajectory-aligned credit allocation to provide fine-grained reward signals for different stages of the search-selection process. Extensive experiments on large-scale tool selection benchmarks demonstrate that ToolSearcher consistently outperforms a set of strong baselines in challenging settings involving iterative search and complex tool composition.

Test-time reinforcement learning (TTRL) has emerged as a promising paradigm for adapting large reasoning models (LRMs) on unlabeled test inputs, using self-consensus rewards derived from majority voting over sampled rollouts. However, majority voting can mistake popularity for correctness: a spurious yet high-frequency unverified consensus may become a biased reward signal, causing test-time RL to reinforce frequent but wrong answers and collapse into an incorrect mode. We address this false-popular failure mode with T^3RL (Tool-Verification for Test-Time Reinforcement Learning), a verification-aware test-time RL framework. T^3RL grounds pseudo-label construction in external tool evidence. Concretely, a verifier uses external tool evidence, such as code execution, to upweight verified rollouts during verification-aware voting, producing more reliable pseudo-labels for training. Across mathematical reasoning benchmarks of varying difficulty, including MATH-500, AMC, and AIME 2024, and diverse backbone families, T^3RL significantly improves over TTRL, with stronger performance on harder problems. More broadly, T^3RL is positioned as a verified online data synthesizer, highlighting the role of test-time tool verification in reliable online adaptation.