NeurIPS 2026 Papers

Decision Transformer (DT) has emerged as a powerful paradigm for offline reinforcement learning. For complex tasks with imbalanced datasets, where high-quality samples constitute only a small fraction of the data while the majority is suboptimal, the performance of DT methods usually degrade significantly due to the the offline data distribution deviates substantially from the stationary distribution of the optimal policy. In this paper, we construct a dual-based correction mechanism to mitigate the distribution shift problem and propose a Distribution Corrected Decision Transformer method. Specifically, we investigate the Lagrangian duality of the reward maximizing objective in the standard Markov decision problem and recover the optimal correction weights between the static offline data and optimal stationary distribution. Note that a double-KL regularizer is integrated in the objective to prioritize high-return state-action pairs while remaining anchored to valid data support. Subsequently, we use the learned weights to correct distributional bias in two critical components of DT: stabilizing policy evaluation through weighted temporal-difference learning, and improving policy extraction by prioritizing DT training on weighted dataset aligned with the corrected target distribution. We theoretically analyze the performance guarantee, proving our method improves upon standard baselines, and empirically demonstrate strong performance on D4RL benchmarks, particularly on highly imbalanced datasets where prior methods fail.

Post-training for reasoning models typically combines supervised fine-tuning with reinforcement learning from verifiable rewards, most commonly with GRPO. However, this algorithm suffers from sparse rewards, limited exploration, and mode collapse. Building upon recent works on self-distillation, we propose Feedback Distillation, a training method where the model is trained to match, at the token level, its own distribution conditioned on privileged feedback produced by a language model. Feedback Distillation offers token-level supervision and can inject external knowledge. Evaluating our method for Lean 4 theorem-proving, we find that Feedback Distillation maintains greater diversity in generated trajectories than GRPO, yielding higher policy entropy and better pass@$k$ scaling. The two methods are complementary: initializing GRPO from a Feedback Distillation checkpoint outperforms either method alone. All in all, our results suggest a promising avenue to improve post-training for complex reasoning.

Current end-to-end autonomous driving models are fundamentally constrained by the behavioral cloning ceiling of imitation learning. While reinforcement learning offers a path to smarter autonomy, it demands two missing pieces of infrastructure: (1) a cognitive foundation that understands traffic semantics and driving intent, and (2) a foresighted physical environment that can anticipate the consequences of candidate actions. To this end, we propose **CoPhy**, a **Co**gnitive-**Phy**sical reinforcement learning framework for autonomous driving. To *distill to think*, we distill VLM knowledge into the BEV encoder and then discard the VLM entirely, retaining cognitive ability at zero inference cost while releasing the cognitive channel as a pluggable interface for optional human language commands. To *foresee to act*, we build an auto-regressive BEV world model that explicitly predicts future semantic maps conditioned on candidate actions, serving as an interpretable physical sandbox from which safety metrics are directly derived. Built upon this dual infrastructure, we optimize the driving policy via GRPO with a novel dual-reward mechanism: a *physical* reward derived from BEV rollouts enforces hard safety constraints, while a *cognitive* reward from a language-aligned scorer ensures intent compliance. Extensive experiments demonstrate that CoPhy not only achieves state-of-the-art results on NAVSIM v1 and v2 benchmarks, but also enables safer driving via cognitively informed scene compliance and flexible intent control through user-defined language instructions.

Model quantization has advanced tremendously in recent years, producing models that require a fraction of the compute while maintaining baseline accuracy and creating the illusion of lossless compression. Beneath the surface, however, model behavior undergoes severe shifts that simple global accuracy fails to capture. Notably, quantized models experience a high rate of "flips": individual predictions that change from correct to incorrect (bad flips-CI) and vice versa (good flips-IC). Although overall accuracy remains stable, these changes degrade model reliability and introduce unforeseen safety risks. Current literature largely attributes these flips to random rounding noise, a perspective that ignores the specific precision constraints imposed by quantization with or without training. In this paper, we show that flips are driven by quantization's implicit regularization, low-pass filtering, and exacerbated spectral bias. By analyzing these effects via Singular Value Decomposition (SVD), we leverage the model's low-to-high rank internal disagreement to disentangle good flips from bad ones. Building on this insight, we introduce post-hoc approaches to selectively preserve the former while correcting the latter, and we validate the universality of our disentanglement through extensive experiments spanning Post-Training Quantization (PTQ) to Quantization-Aware Training (QAT) and across vision, language, and multi-modal models. With this work, we aim not only to restore the quantized model's hidden reliability degradation but also to elevate it beyond the accuracy ceiling of its full-precision counterpart.

Masked diffusion language models generate text through iterative masked-token filling, but terminal-only rewards on final completions provide coarse credit assignment for the intermediate filling decisions that shape the generation process. We propose Diffusion-State Policy Optimization (DiSPO), a plug-in credit-assignment layer that directly optimizes intermediate filling decisions. At selected intermediate masked states, DiSPO branches by resampling the currently masked positions from rollout-cached logits, scores the resulting completions, and updates only the newly filled tokens, requiring no additional multi-step diffusion rollouts or optimizer steps. We formalize a fixed-state objective for branched completions and derive a policy-gradient estimator that reuses the same rollouts as terminal-feedback policy optimization. Experiments on LLaDA-8B-Instruct show that DiSPO consistently improves terminal-feedback baselines, including diffu-GRPO and SPG, on math and planning benchmarks under matched rollout compute and optimizer steps, supporting its use as a general plug-in for masked diffusion policy optimization.

Distillation and Reinforcement Learning (RL) fine-tuning are the primary pillars of diffusion post-training. While traditionally studied in isolation, the interaction between these phases remains poorly understood, and in particular how fine-tuning impacts the generative quality of distilled models. We introduce Rewarded Moment Matching Distillation (RMMD), a novel framework that simultaneously distills diffusion models and maximizes a reward function. RMMD preserves the high-fidelity "naturalness" characteristic of advanced distillation (e.g., 8-step Moment Matching) by adapting the sampling loop for on-policy training and repurposing the distillation loss as a proxy for integral KL regularization. By evaluating the FID-Reward Pareto fronts on ImageNet, we demonstrate that RMMD achieves superior trade-offs compared to single-step baselines (DI++) and multi-step competitors (DRaFT, HyperNoise). Finally, we apply RMMD to GenCast, a state-of-the-art weather forecasting model, to distill it while optimizing the Continuous Ranked Probability Score (CRPS) metric. The resulting distilled model achieves a 7.5x speedup while outperforming the teacher model on 93% of target weather variables, and being better calibrated. This proves that RMMD scales to complex, high-dimensional scientific domains.

Offline goal-conditioned reinforcement learning (GCRL) learns goal-directed policies from reward-free data, but in long-horizon tasks, goal-conditioned value functions often provide unstable guidance due to sparse rewards and discounting. Hierarchical methods partially mitigate this issue via subgoal decomposition; however, high-level decision-making still relies on noise-sensitive value estimates, leading to unstable behavior in complex environments. We address this limitation by proposing \textbf{D}iffusion \textbf{S}ubgoal \textbf{P}lanning (\textbf{DSP}), a diffusion-based framework for high-level subgoal generation. DSP casts high-level planning as guided generative inference over goal-conditioned subgoals and learns both conditional and unconditional flows, enabling classifier-free guidance to introduce a goal-directed bias at inference time. By removing explicit value-based guidance from high-level planning, DSP generates reachable and goal-directed subgoals through a generative model while retaining hierarchical execution. Experiments on offline GCRL benchmarks demonstrate that DSP outperforms prior methods on a range of navigation and manipulation tasks, with particularly strong performance in maze environments that require multi-step subgoal planning.

Human coordination often relies on the ability to influence the beliefs of others through strategic action. In multi-agent reinforcement learning, opponent shaping attempts to replicate this influence, though existing methods typically operate within an opponent's parameter, policy, or value space. Meanwhile, belief-manipulation techniques in hidden-role games often rely on hard-coded objectives, such as deception or belief saturation. We propose Differentiable Belief-based Opponent Shaping (D-BOS), a first-order method that treats each observer's belief as the shaped opponent state and differentiates through $k$-step softmax-Bayes belief dynamics. Rather than explicitly rewarding deceptive or cooperative behavior, our method treats the belief state as the target for shaping. This allows the optimal strategy to emerge naturally from the environment's reward structure. This belief-space formulation provides an opponent-shaping signal by differentiating through opponent belief updates, and naturally extends to multiple observers by aggregating gradients over their individual inferred belief trajectories. Empirically, D-BOS outperforms PPO and BBM in hidden-role games, with the largest gains in mixed-motive settings.

In reinforcement learning (RL), reward specification plays the fundamental role in shaping policy optimization. Inverse RL (IRL) offers a data-driven approach to inferring rewards from demonstrations, but prevalent neural-network-based methods often produce rewards lacking interpretability and flexibility for manual inspection or revision. In contrast, human-written or large language model (LLM)-generated reward code provides explicit semantic structure, yet its precise numerical parameters are often difficult to calibrate and prone to specification errors. To bridge this gap, we propose DiReCL (\textbf{Di}fferentiable \textbf{Re}ward \textbf{C}ode \textbf{L}earning), a novel framework that unifies these two paradigms by formulating reward learning in a differentiable reward code space. DiReCL first prompts an LLM to synthesize the reward code templates, then converts the numeric elements in these templates into differentiable parameters and align them with expert demonstrations through IRL. We further introduce a reflection mechanism that uses IRL diagnostics to provide evidence of inadequacies in the reward structure and guide template revision. Extensive experiments on MuJoCo locomotion and autonomous driving benchmarks demonstrate that DiReCL achieves over $1.25 \times$ the performance of both standard IRL and LLM-assisted reward design baselines, while preserving the crucial benefits of explicit, interpretable reward semantics.

Dexterous object manipulation remains a fundamental challenge in robotics, typically requiring intensive supervision through explicit rewards, goals, or demonstrations. While unsupervised reinforcement learning (RL) holds the promise of discovering dexterous skills, state-of-the-art metric-aware methods have yet to be successfully applied to complex, high-dimensional manipulators like multi-fingered hands. We argue that this limitation stems from two key factors: existing methods lack the necessary inductive biases to prioritize object interaction over agent motion, and they fail to respect the non-Euclidean topology of object configuration space $\mathrm{SE}(3)$. To address these limitations, we propose a novel Wasserstein dependency measure between skills and object motions, formulated in the Lie algebra $\mathfrak{se}(3)$. Our approach, Topology-Aware SKill discovery (TASK), enables the discovery of fundamental manipulation skills—such as multi-axial object translation and rotation—without manual supervision. Across a range of embodiments, including non-prehensile and multi-fingered hands, our method learns dexterous behaviors where previous approaches fail, marking a significant step toward autonomous robotic dexterity.

Reinforcement learning (RL) post-training provides a direct way to align diffusion and flow-matching generators with human preferences and task-specific rewards, but current RL algorithms for diffusion models remain fragmented: reverse-trajectory methods rely on discretized likelihood ratios, whereas forward-matching methods train on reward-labeled noising pairs. This paper shows that these seemingly different losses arise from a single path-space principle. Starting from the regularized diffusion-RL objective, we use importance sampling between sampling SDEs to obtain an explicit policy-gradient estimator on trajectory space. The estimator contains the stochastic It\^o integral underlying Flow-GRPO-type updates; we derive an equivalent variance-reduced value-gradient form that recovers the forward-matching structure of AWM and DiffusionNFT. This identifies the empirical gap between these method families as a variance-reduction effect rather than a difference in RL principle. The derivation yields a unified design space organized by value-gradient estimation, weight functions, and sampling choices. Within this space, we propose a multi-sample KDE value-gradient estimator that reuses rollout groups, together with scale-bounded weight families that retain stable existing recipes while excluding singular ones. Experiments on SD3.5-M and Qwen-Image models validate the variance-reduction explanation and show that the resulting recipe matches or improves over prior diffusion-RL baselines.

Reinforcement learning (RL) has become a prevalent paradigm for training tool calling agents, which typically requires online interactive environments. Existing approaches either rely on training data with ground truth annotations or require advanced proprietary language models (LMs) to synthesize environments that keep fixed once created. In this work, we propose TRUSTEE, a cost-friendly method for training tool calling agents with dynamic environments fully simulated by free open-source LMs that can be as small as 8B, including task generation, user simulation, tool simulation and trajectory evaluation, paired with an adaptive curriculum learning mechanism that controls task difficulty during training. Our empirical results show that TRUSTEE outperforms baselines which require extra external resources in most cases. These confirm that, with a sufficiently sophisticated design, even simulated environments with a local 8B LM as the backbone could set a strong baseline for tool learning. We hope our proposed paradigm could democratize tool learning and inspire future research on environment scaling with limited resources.

Distributed reinforcement learning trains on data from stale, buggy, or mismatched actors, producing actions with high surprisal (negative log-probability) under the learner's policy. The core difficulty is not surprising data per se, but \emph{negative learning from surprising data}. High-surprisal failures can dominate finite-batch updates through large perpendicular components, while high-surprisal successes reveal opportunities the current policy would otherwise miss. The \textit{Delightful Policy Gradient} (DG) separates these cases by gating each update with delight, the product of advantage and surprisal, suppressing rare failures and preserving rare successes without behavior probabilities. In a tabular analysis, DG suppresses the perpendicular second moment of high-surprisal failures by a policy-overlap factor that vanishes as the learner improves. The advantage sign is essential for surprisal-based filtering: any learner-probability-only gate that suppresses rare failures also suppresses rare successes. On MNIST with simulated staleness, DG without off-policy correction outperforms importance-weighted PG with exact behavior probabilities. On a transformer sequence task with staleness, actor bugs, reward corruption, and rare discovery, DG often achieves nearly order-of-magnitude lower error. When all four frictions act simultaneously, its sample-efficiency advantage is order-of-magnitude and grows with task complexity.

Uncertainty quantification (UQ) in deep learning regression is of wide interest, as it supports critical applications including sequential decision making and risk-sensitive tasks. In heteroskedastic regression, where the uncertainty of the target depends on the input, a common approach is to train a neural network that parametrises the mean and the variance of the predictive distribution. Yet to this day, training deep heteroskedastic regression models poses severe practical challenges in the trade-off between uncertainty quantification and mean prediction, such as variance overfitting, optimization difficulties and representation collapse. In this work, we identify the core issues and propose a simple and efficient procedure that addresses these jointly by post-hoc fitting a variance model across the intermediate layers of a pretrained network on a hold-out dataset. We demonstrate that this method is competitive with end-to-end trained mean-variance networks in heteroskedastic UQ on several data modalities. The method retains mean prediction accuracy, is cheap at both train and prediction time, and requires no additional data compared to existing methods. Finally, we show this method works on large scale foundation models for chemistry, paving the way for cheap heteroskedastic UQ in large scale neural networks without having to retrain such models end-to-end.

Double Q-learning is a classical control algorithm that mitigates the maximization bias of Q-learning. To do so, it explicitly trains two independent action-value functions and uses them to decouple action-selection and action-evaluation when computing bootstrap targets. Double DQN adapts target bootstrap decoupling to deep reinforcement learning (RL), but explicitly trains only a single action-value function and does not fully decouple its estimators. Consequently, the two estimators remain correlated, and overestimation persists. In this paper, we introduce Deep Double Q-learning (DDQL), a deep RL algorithm that explicitly trains two Q-functions through Double Q-learning. DDQL stabilizes training through a combination of techniques, including lower replay ratios, longer target network update intervals, and shared layers. Across 57 Atari 2600 games, DDQL improves aggregate performance over Double DQN, outperforming it on 47 games while further reducing overestimation. In addition, we study key design choices when adapting Double Q-learning to deep RL, including the network architecture, replay ratio, and minibatch sampling strategies.

The process of discovery requires active exploration---the act of collecting new and informative data. However, efficient autonomous exploration remains a major unsolved problem. The dominant paradigm addresses this challenge by using Reinforcement Learning (RL) to train agents with intrinsic motivation, maximizing a composite objective of extrinsic and intrinsic rewards. We suggest that this approach incurs unnecessary overhead: while policy optimization is necessary for precise task execution, employing such machinery solely to expand state coverage may be inefficient. In this paper, we propose a new approach that explicitly decouples exploration from policy optimization and bypasses RL entirely during the exploration phase. Our method uses a tree-search strategy inspired by the Go-With-The-Winner algorithm, paired with a measure of uncertainty to systematically drive exploration. By removing the overhead of policy optimization, our approach explores an order of magnitude more efficiently than standard intrinsic motivation baselines on hard exploration benchmarks. Further, we demonstrate that the trajectories discovered during exploration can be distilled into deployable policies using existing supervised backward learning algorithms, achieving state-of-the-art performance by a wide margin on Montezuma's Revenge, Pitfall!, and Venture without relying on domain-specific knowledge. Finally, we demonstrate the generality of our framework in high-dimensional continuous action spaces by solving the MuJoCo Adroit dexterous manipulation and AntMaze tasks in a \emph{sparse-reward} setting, directly from image observations and without expert demonstrations or offline datasets. To the best of our knowledge, this has not been achieved before.

Recent full-duplex spoken dialogue models have demonstrated compelling progress toward human-like interaction, enabling agents to respond with low latency, produce backchannels, and handle user barge-ins. Yet these improvements in conversational dynamics often come with weaker reasoning and instruction-following abilities, revealing a potential tension between interactive dynamics and intelligence capability. In this paper, we argue that such an **intelligence--dynamics** trade-off is not fundamental: conversational dynamics can instead be learned as a separate real-time decision policy from human dialogue data. To this end, we propose DuplexPO, a reinforcement learning (RL) framework that decouples when to speak from what to say. It preserves the semantic response capability of an instruction-tuned assistant, while optimizing its temporal interaction behavior over selected high-impact windows from long human conversations. To quantitatively optimize these dynamics, we formulate the Factorized Conversational Dynamics Reward (FCDR) to enable fine-grained temporal credit assignment for turn initiation, backchanneling, yielding, and regularized participation. The policy is then optimized with a GRPO-style objective. Experiments show that DuplexPO substantially improves full-duplex behaviors, including timely backchannels, smooth turn-taking, and barge-in handling, while maintaining strong reasoning and instruction-following performance. Moreover, improvements in dynamics-oriented metrics are reflected in better user experience, suggesting that optimizing conversational timing as a standalone objective can promote more natural full-duplex interaction.

Identifying subpopulations of neurons that interact with each other from simultaneous recordings of populations of many neurons is key for understanding across-brain communication with cellular resolution. Recent work identified communication subspaces, which capture additive interactions between pairs of high-dimensional neural populations through a small number of source and target activity patterns. However, no current method captures how a third, potentially multivariate variable - such as behavioral state or the activity of a third population - modulates these interactions. Here we extend the communication subspace framework by parameterizing modulation as a low-rank tensor. This identifies multiplicative interaction channels (MICs), defined as triplets of source, target, and modulator activity patterns, in which the modulator pattern gates the source-target interaction. We derive MICs as a bilinear perturbation of reduced-rank regression. We develop a hierarchical fitting pipeline and provide a closed-form decomposition that quantifies whether modulation reshapes the modulator-averaged baseline interaction, recruits private dimensions of one population, or opens new interactions. In simulations, MICs reliably recover the presence and geometry of ground-truth modulation even in the high-dimensional, low-sample regime. Applying MICs to simultaneous calcium imaging of prefrontal axons and interneurons in the visual cortex revealed that behavioral state asymmetrically modulates top-down interactions, reconfiguring the patterns of prefrontal projections that interact with a stable set of visual interneuron activity patterns. By providing an efficient and compact characterization of modulatory interactions, MICs enable asking new questions about how potentially high-dimensional variables shape interactions between neural populations.

Generative visual verifiers play a crucial role in advancing multimodal intelligence systems. In this paper, to enhance the robustness of visual verifiers in complex and realistic open-world scenarios, we move beyond conventional binary verifiers that produce only True/False discrete judgments. Instead, we introduce a verifier that automatically decomposes complex prompts or environmental rules into fine-grained components and assigns pointwise scores, thereby providing a continuous and more informative critique signal. Under a general reinforcement learning framework, we first empirically demonstrate that the proposed pointwise decomposed verifier significantly outperforms binary verifiers in traditional text-to-image scenarios. We then extend our investigation to a novel paradigm: generative world modeling, where generative models are leveraged for world simulation and world reasoning. From a theoretical standpoint, we further show that when ground-truth images are easy to obtain, a pairwise verification paradigm yields more accurate critiques than the pointwise formulation. Empirical results on diverse world-modeling tasks, such as Maze and Sudoku, further validate the effectiveness of our approach. Overall, our findings suggest a key insight: the transition from generative models to world-modeling agents critically hinges on the availability of accurate pairwise decomposed visual verifiers.