NeurIPS 2026 Papers

In-context reinforcement learning (ICRL) leverages the in-context learning capabilities of transformer models (TMs) to efficiently generalize to unseen sequential decision-making tasks without parameter updates. However, existing ICRL methods rely on explicit reward signals during pretraining, which limits their applicability when rewards are ambiguous, hard to specify, or costly to obtain. To overcome this limitation, we propose a new learning paradigm, \emph{In-Context Preference-based Reinforcement Learning} (ICPRL), in which both pretraining and deployment rely solely on preference feedback, eliminating the need for reward supervision. We study two variants that differ in the granularity of feedback: \emph{Immediate Preference-based RL} (I-PRL) with per-step preferences that generalize dueling bandits, and \emph{Trajectory Preference-based RL} (T-PRL) with trajectory-level comparisons. We first show that supervised pretraining, a standard approach in ICRL, remains effective under preference-only context datasets, demonstrating the feasibility of in-context solving unseen tasks using only preference signals. To further improve data efficiency, we introduce alternative preference-native frameworks for I-PRL and T-PRL that directly optimize TM policies from preference data without requiring reward signals nor optimal action labels. Experiments on dueling bandits, navigation, and continuous control tasks demonstrate that ICPRL enables strong in-context generalization to unseen tasks, achieving performance comparable to ICRL methods trained with full reward supervision.

As agents have to solve diverse tasks from grid-worlds to robotics and language modeling, the difficulty of each step varies, yet policies are rigid and spend a fixed amount of compute on each step. A natural approach is to spend compute on the harder states while saving it on easy ones, allowing agents to retain performance while saving valuable compute budgets. To this end, we propose Adaptive Internal Computation (AIC), a lightweight policy architecture that iteratively refines latent states and learns when to halt using a value-of-computation signal.Concretely, AIC repeatedly refines a latent state representation and learns a halting head from a value-of-computation signal, the per-step gain in expected return, allowing the policy to stop refining as soon as further computation no longer improves the action it would take. A central challenge, however, is that matched-budget gains alone cannot establish that compute is being directed to the states that benefit the most from them. We therefore introduce a six-component causal evaluation protocol leveraging a targeted-versus-random ablation that tests whether removing compute from AIC's predicted high-compute states hurts performance more than removing the same amount of compute from randomly chosen states, thereby validating that AIC can successfully identify high-compute states. Across twelve diverse tasks, including grid worlds, simulated robotics experiments, and language modeling, we demonstrate that AIC outperforms fixed-compute policies across a set of five underlying model architectures. In particular, architectures leveraging AIC retain or improve performance while spending about 52\% less compute compared to their respective fixed counterparts.

Behavior cloning provides strong imitation learning guarantees when training and test environments share the same dynamics. However, in many deployment settings the test environment's transitions differ from training, and classical offline IL offers no recourse: the learner must commit to an action at every state, even when its demonstrations are uninformative and could lead to arbitrary degradation of performance. This motivates the study of _selective_ imitation, where the learner may choose to _stop_ when it cannot act reliably. We introduce a model for selective imitation under arbitrary dynamics shift: given labeled expert demonstrations from a training environment and unlabeled state trajectories from the same expert in a test environment, the learner outputs a _selective policy_ that is _complete_ (rarely stops in training) and _sound_ (incurs low regret before stopping in test). Our algorithm, $\operatorname{SeqRejectron}$, constructs a stopping rule using a small set of _validator policies_ whose size is independent of the horizon or policy class. For deterministic policies, this yields horizon-free $\tilde{O}(\log|\Pi|/\epsilon^2)$ sample complexity, assuming sparse costs. For stochastic policies, we obtain analogous horizon-free guarantees using a cumulative Hellinger stopping time. We extend the framework to misspecified experts and different expert policies across train and test and obtain results that gracefully degrade with the amount of misspecification.

Whether navigating a building, operating a robot, or playing a game, an agent that acts effectively in an environment must first learn an internal model of how that environment works. Partially observable Markov decision processes (POMDPs) provide a flexible modeling class for such internal world models, but learning them from observation-action trajectories alone is challenging and typically requires extensive environment interaction. We ask whether language-model priors can reduce costly interaction by leveraging prior knowledge, and introduce Pinductor (POMDP-inductor): an LLM proposes candidate POMDP models from a few observation-action trajectories and iteratively refines them to optimize a belief-based likelihood score. Despite using strictly less information, Pinductor matches the performance and sample efficiency of LLM-based POMDP learning methods that assume privileged access to the hidden state, while significantly surpassing the sample efficiency of tabular POMDP baselines. Further results show that performance scales with LLM capability and degrades gracefully as semantic information about the environment is withheld. Together, these results position language-model priors as a practical tool for sample-efficient world-model learning under partial observability, and a step toward generalist agents in real-world environments.

Value overestimation is a central challenge in offline reinforcement learning (RL), arising when Bellman backups propagate out-of-distribution (OOD) errors that lead to catastrophic inflation of $Q$-values. Prior approaches address this issue by encouraging the learned policy to stay close to the behavior policy, but they often result in suboptimal policies or require significant training complexity. More recent work attempts to tackle this issue through minimal modifications to the critic network, yet it remains limited in either theoretical grounding or practical implementation. In this work, we present Kernel Value Regression (KVR), a simple critic regularization method based on kernel regression. KVR builds on the vanishing extrapolation property of kernel ridge regression, which encourages value estimates to decay toward zero away from the data support. This inductive bias naturally suppresses $Q$-values for OOD actions without auxiliary regularizers, providing a structural mechanism for addressing value overestimation. Using random Fourier features, KVR can be incorporated into standard actor-critic algorithms with only minimal modifications. We empirically demonstrate that these properties persist in complex offline RL environments, leading to strong performance across diverse tasks in the D4RL and OGBench benchmarks.

Meta-learned intrinsic rewards are a powerful tool for shaping policy learning in reinforcement learning, particularly when extrinsic rewards are sparse, delayed, or unavailable. Learning Intrinsic Rewards for Policy Gradients (LIRPG) introduced this paradigm and serves as the foundation on which subsequent meta-learned intrinsic reward methods are built. While LIRPG and LIRPG-based methods have shown strong results in low-dimensional control, their behavior in high-dimensional domains, where the policy and value networks use a shared encoder, remains poorly understood. In this work, we analyze meta-learned intrinsic rewards in high-dimensional environments and uncover a consistent failure mode, particularly when training with intrinsic rewards alone, where performance collapses to near-random behavior. We identify the root cause as dense, non-stationary intrinsic rewards inducing large and high-variance value losses that dominate shared encoder updates, suppressing policy learning. We further demonstrate that decoupling policy and value optimization using phasic policy gradient methods is one simple and effective approach to addressing this issue.

We propose a novel definition of model exploitation in reinforcement learning. Informally, a world model is exploitable if it implies that one policy should be strictly preferred over another while the environment’s true transition model implies the reverse. We analogize our definition with a prior characterization of reward hacking but show that the associated proof of inevitability does not transfer to exploitation. To overcome this obstruction, we develop a general theory of reward hacking and model exploitation that proves that exploitation is essentially unavoidable on large policy sets and yields the corresponding claim for hacking as a special case. Unfortunately, we also find that the conditions that guarantee unhackability in finite policy sets have no counterparts that preclude exploitation. Consequently, we introduce a relaxed notion of exploitation and derive a safe horizon within which it can be avoided. Taken together, our results establish a formal bridge between reward hacking and model exploitation and elucidate the limits of safe planning in world models.

In-context reinforcement learning (ICRL) is often described as inference-time RL in which an LLM improves by accumulating trajectory-reward pairs in its context, with the reward acting as a learning signal. This framing poses a central question: is ICRL genuine inference-time RL, or better understood as a form of in-context learning (ICL)? We investigate this in the context of direct ICRL, where the model directly uses trajectory-reward pairs. Through controlled experiments on three benchmarks across six models, we find that the reward is read, but its effect is small; randomizing or removing the reward leaves the improvement curve almost unchanged. Trajectories drive improvement, but not through their semantic content: shuffled or corrupted trajectories work as well as real ones. These patterns closely mirror those known in ICL, suggesting that direct ICRL is better understood as a special case of ICL than as inference-time RL. This reframing has implications for ICRL memory design: ICL factors such as input distribution and demonstrations may matter more than RL elements such as reward shaping and exploration.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a promising approach to enhance Instruction Following (IF) capabilities of large language models (LLMs). However, RLVR for instruction following remains sample-inefficient and prone to reward hacking, where LLMs exploit verification shortcuts rather than fulfilling the core intent. To address these challenges, we frame RLVR for instruction following as an integrated environment that unifies dynamic task generation and robust training. We introduce Instruction Following Decorator (IFDecorator}, a framework coupling difficulty adaptation, intent alignment, and hack diagnostics. It features: (1) a cooperative-adversarial data flywheel that yields challenging yet solvable tasks for sample-efficient training; (2) IntentCheck, a gating module to mitigate reward hacking; and (3) TripWires, a proactive diagnostic tool for eliciting hacking behaviors and quantifying their prevalence. Extensive experiments show our Qwen2.5-32B-Instruct trained with IFDecorator using only 3,625 training examples achieves 87.43% on IFEval (outperforming GPT-4o) and improves FollowBench by 4.2%, while preserving general capabilities. Diagnostics confirm that our method effectively reduces reward hacking. The approach generalizes across model architectures and scales. We will release code and data for future research.

High-dimensional robotic reinforcement learning (RL) is often bottlenecked by inefficient exploration in large, redundant actuator spaces. We introduce Devyn (Developing synergies), a lightweight framework that learns a structured low-dimensional action parameterization directly during policy optimization. Unlike methods that require precomputed synergies or structured exploration in the full action space, Devyn starts from a randomly initialized latent-to-action map and shapes it online through task interaction. The policy acts strictly in a reduced latent space, while a learnable synergy matrix decodes latent actions into actuator commands. To make this evolving decoder stable and useful for policy learning, Devyn combines slow decoder evolution with structural regularization, reducing latent-action semantic drift while encouraging compact coordination patterns. Across diverse continuous-control benchmarks, including high-dimensional humanoid and overactuated musculoskeletal systems, Devyn improves upon standard SAC and achieves competitive or better performance than recent methods designed for high-dimensional robotic control. Ablations show that the gains do not come from a generic latent-action bottleneck alone, but rather from the slow, structured evolution of the synergy matrix. Our results suggest that learning the action parameterization itself can be a simple and effective route to scaling RL for high-dimensional robots.

We present, to our knowledge, the first end-to-end FP4 RL post-training, in which both the rollout and training policies, including their forward and backward passes, operate at 4-bit precision. A systematic study reveals that the dominant source of degradation in FP4 RL is not training-side quantization error but rollout activation quantization: outliers stretch the dynamic range so far that a large number of activation values underflow to zero under FP4. Counterintuitively, restoring the training policy to higher precision while keeping the rollout in FP4 makes accuracy worse than full FP4 baseline, exposing rollout–training mismatch as the principal failure mode and ruling out standard pretraining-style fixes. We address this with Rollout Residual Quantization (Rollout-ResQ): a single residual correction term constrained to a hardware-friendly sparsity pattern, added only to the FP4 rollout matmul — a lightweight correction that recovers most of the precision lost to outlier-driven underflow without inflating the rollout’s compute footprint. On Qwen2.5-3B and Qwen2.5-Math-7B, Rollout-ResQ paired with the HiFloat4 (HiF4) format — whose three-level hierarchical scaling preserves resolution under FP4’s tight 4-bit budget — closes the accuracy gap to BF16 from 4.9% to 1.1%, bringing fully quantized FP4 RL within striking distance of full precision. Applied to the open-standard MXFP4, the same recipe narrows the gap from 13.6% to 5.3%, revealing that FP4 format choice is a key factor that determines the ceiling on recoverable accuracy. Together, these results establish HiF4 as the enabling format for end-to- end FP4 RL post-training, and Rollout-ResQ as the activation-side mechanism that makes the gap to BF16 closable.

Prioritized Sweeping (PS) accelerates model-based reinforcement learning by selecting backups according to residual magnitude. In nonstationary reward settings, however, the canonical priority score is shortsighted: after a localized reward shift, residuals propagate only through realized backups, so bottlenecked or topologically distant state estimates may remain static under a limited replanning budget. We introduce Graph Topology Augmentation for Prioritized Sweeping (GTA-PS), a concrete instance of a broader topology-augmentation principle for priority-based planning. GTA-PS constructs a policy-induced transition graph and augments the residual key with a mixing of regularized directed-Laplacian potentials that diffuses residual information through forward and backward graph structure. The topology weight is controlled by a scheduler based on the Second Largest Eigenvalue Modulus (SLEM), allowing the queue to adapt to the chain's mixing regime. We prove that the forward potential coincides with discounted residual propagation at a canonical regularization parameter and show that GTA-PS gives nonzero priority to states that standard PS can leave blocked after sparse reward shifts. Tabular experiments on FourRooms and GARNET domains demonstrate improved replanning efficiency over standard PS under both exact DP and Dyna-style host planners.

Diffusion distillation, exemplified by Distribution Matching Distillation (DMD), has shown great promise in few-step generation but often sacrifices quality for sampling speed. While integrating Reinforcement Learning (RL) into distillation offers potential, a naive fusion of these two objectives relies on suboptimal raw sample evaluation. This sample-based scoring creates inherent conflicts with the distillation trajectory and produces unreliable rewards due to the noisy nature of early-stage generation. To overcome these limitations, we propose GDMD, a novel framework that redefines the reward mechanism by prioritizing distillation gradients over raw pixel outputs as the primary signal for optimization. By reinterpreting the DMD gradients as implicit target tensors, our framework enables existing reward models to directly evaluate the quality of distillation updates. This gradient-level guidance functions as an adaptive weighting that synchronizes the RL policy with the distillation objective, effectively neutralizing optimization divergence. Empirical results show that GDMD sets a new SOTA for few-step generation. Specifically, our 4-NFE (Number of Function Evaluations) models outperform the quality of their multi-step teacher and substantially exceed previous DMDR results in GenEval and human-preference metrics, exhibiting strong scalability potential.

The open-ended generation in LLMs usually requires multi-dimensional rubrics to adequately assess quality and guide the improvement of reinforcement learning. However, a critical dilemma inherent in this training paradigm is the imbalanced reward polarization along different rubric dimensions. Under this bottleneck, even if LLMs achieve relatively high rewards after training, they may still exhibit severe deficiencies in certain dimensions, leading to a direct deterioration in user experience. To address this problem, we propose \emph{Focal Reward}, a novel objective to automatically balance the training of reinforcement learning under rubric-based rewards. Specifically, we first leverage an inverse reward projection mechanism to estimate the saturation degree of each criterion in the rubric, which forms the basis to calibrate the reward direction. Then, the final objective is designed with an automatically reweighting coefficient for each criterion to achieve the fine-grained balancing. Extensive experiments across three model scales and six benchmarks demonstrate that our \emph{Focal Reward} method outperforms the strongest static aggregation baseline in all 18 model--benchmark comparisons. Rollout, mechanism, and ablation analyses further show that these gains arise from online, saturation-aware reallocation toward rubrics that still have room for improvement.

We develop a unified framework for flow-based reinforcement learning (RL) grounded in diffusion–flow duality and establish its theoretical foundations. The frameworks enables RL for deterministic flow matching inference by introducing diffusion stochastic inference dynamics that supports exploration while admitting deterministic deployment. We show that our frameworks subsumes several existing flow-based RL methods as special cases and inspire effective now methods. Then we provide theoretical foundations for the frameworks by establishing the correctness and performance transfer guarantee. Specifically, we prove that the policies optimized under stochastic dynamics close to deterministic dynamics at deployment when the pretrained model is well-trained. Moreover, we show that the policy improvement achieved under training-time SDE inference transfers to deployment-time ODE inference. Finally, we conduct experiments to validate our theoretical results.

Building agentic systems that can autonomously self-improve from experience is a longstanding goal of AI. Large language models (LLMs) today primarily self-improve via two mechanisms: self-reflection for context updates, and reinforcement learning (RL) for weight updates. In this work, we propose Evolutionary System Prompt Learning (E-SPL), a method for jointly improving model contexts and model weights. In each RL iteration, E-SPL samples trajectories under multiple system prompts in parallel, then jointly applies RL updates to weights and evolutionary updates to system prompts via LLM self-reflection. E-SPL encourages a natural division between declarative knowledge encoded in prompts and procedural knowledge encoded in weights. Most notably, in an easy-to-hard generalization setting (AIME $\rightarrow$ BeyondAIME), E-SPL improves RL success rate from 38.8% $\rightarrow$ 45.1%. E-SPL also improves RL on AIME 2025 (56.3% $\rightarrow$ 60.6%), HMMT 2025 (50.0% $\rightarrow$ 52.7%), and agentic search (44.2% $\rightarrow$ 48.6% on gpt-oss-120b). Across all settings, E-SPL outperforms both RL-only and evolution-only baselines, demonstrating that weight updates and context updates are deeply synergistic and can together yield gains in generalization that neither achieves alone.

Recent multimodal large language models (MLLMs) achieve strong results on video question answering, yet often fail to recover the temporal structure that makes a video coherent. They may merge neighboring events, miss single-event progression, or infer implausible relations across events. We study structured temporal understanding as the ability to recover event structure from visual evidence, including event boundaries, single-event progression, and multi-event relations. We argue that this gap arises from a mismatch between current post-training objectives and temporal reasoning, where caption-style supervision allows models to rely on language shortcuts rather than visual temporal grounding. To address this, we propose $\textbf{EventLens}$, a vision-centric reinforcement learning framework that learns video understanding through event-structure recovery}. Instead of reconstructing captions, EventLens trains models to recover temporal structure under controlled perturbations, such as altered temporal granularity, reversed progression, and shuffled event order. We instantiate this framework with a three-level temporal ontology and derive three verifiable task families: event segmentation, progression discrimination, and multi-event relation reasoning. These tasks admit deterministic rewards, enabling task-specialized GRPO training without human preference labels or learned reward models. We further introduce multi-teacher on-policy distillation (MT-OPD) to consolidate specialized policies into a unified model. Experiments on Qwen3-VL backbones show consistent improvements on temporal grounding and event-sensitive reasoning benchmarks, while preserving general video-QA performance. Learning curves further demonstrate that EventLens achieves comparable downstream transfer with significantly reduced training cost compared to direct mixed RL optimization.

Offline goal-conditioned reinforcement learning remains challenging in stochastic long-horizon settings, where compounding value estimation errors hinder reliable goal reaching. While prior methods have sought to address this challenge through various approaches, a key challenge remains path mirage, where agents overcommit to spuriously successful trajectories in offline datasets induced by stochasticity, often leading to failure on difficult tasks. To address this issue, we propose Branch-aware Graph Planning (BGP), which captures stochastic branching structures by identifying high-variability points as graph nodes and designing goal-conditioned edges with segmented learning to avoid unreliable branches and reduce unnecessary stochasticity along paths. As a result, BGP enables reliable goal reaching even in highly stochastic long-horizon environments, and experiments on diverse OGBench tasks show that it substantially outperforms prior state-of-the-art offline goal-conditioned RL methods.

We consider the problem of synthesizing Clifford quantum circuits for devices with all-to-all qubit connectivity. We approach this task as a reinforcement learning problem in which an agent learns to discover a sequence of elementary Clifford gates that reduces a given symplectic matrix representation of a Clifford circuit to the identity. This formulation permits a simple learning curriculum based on random walks from the identity. We introduce a novel neural network architecture that is equivariant to qubit relabelings of the symplectic matrix representation, and which is size-agnostic, allowing a single learned policy to be applied across different qubit counts without circuit splicing or network reparameterization. On six-qubit Clifford circuits, the largest regime for which optimal references are available, our agent finds circuits within one two-qubit gate of optimality in milliseconds per instance, and finds optimal circuits in 99.2\% of instances within seconds per instance. After continued training on ten-qubit instances, the agent scales to unseen Clifford tableaus with up to thirty qubits, including targets generated from circuits with over a thousand Clifford gates, where it achieves lower average two-qubit gate counts than Qiskit's Aaronson-Gottesman and greedy Clifford synthesizers.