NeurIPS 2026 Papers

Reward-robust RL typically models misspecification by specifying an uncertainty set that constrains the discrepancy of plausible rewards from a base reward. However, such reward space discrepancies can be behaviorally irrelevant: under potential-based reward shaping (PBRS), many distinct rewards preserve policy ordering. This can introduce substantial redundancy into standard uncertainty sets and degrade optimization performance. We propose **Shaping-Aware Reward-Robust RL**, which constructs uncertainty sets over PBRS equivalence classes by projecting each reward to a canonical representative, ensuring that the resulting set contains only rewards that induce behaviorally distinct policy rankings. We prove that this projection preserves the optimal robust value while shrinking the uncertainty set and improving empirical performance. Using the connection between robustness and regularization, we obtain a practical algorithm to solve the shaping-aware reward-robust RL problem and enjoy convergence guarantees under standard assumptions. Experiments on benchmarks spanning diverse task domains and levels of complexity show consistent improvements over representative robust RL baselines and exhibit improved robustness to reward perturbations.

Reinforcement learning (RL) is increasingly used to improve the reasoning, coding, and tool-use capabilities of large language models, but agentic RL remains prohibitively expensive. Scaling RL to agentic LLMs requires supporting complex workloads, including multi-policy collaborative training, while efficiently using elastic, heterogeneous, and cross-region compute resources. Existing LLM RL systems support some of these capabilities, but each new extension often requires dedicated system engineering. This burden arises from trainer-centered control architectures and the lack of principled abstractions for RL system components. To address these limitations, we propose AstraFlow, a dataflow-oriented RL system that replaces conventional trainer-centered control with principled component abstractions. In AstraFlow, rollout services, dataflow management, and training are decoupled into autonomous components, enabling the system to natively support complex multi-policy agentic RL workloads and efficiently exploit diverse compute resources. We evaluate AstraFlow across math, code, search, and AgentBench workloads, showing that the same system supports multi-policy training, elastic scaling, heterogeneous cross-region execution, and composable data algorithms without system-level code changes. In multi-policy collaborative training, AstraFlow achieves comparable or better accuracy than existing RL systems while speeding up training time by 2.7x.

Temporally extended exploration via graph Laplacian-based options is a promising approach to sparse-reward reinforcement learning (RL), but existing methods either do not explicitly target novelty or fail to scale to pixel-based domains under function approximation. Novel Exploration via Orthogonality (NEO) addresses the first issue by constructing options that navigate from highly visited regions toward less visited ones, yet prior results were limited to settings where exact eigenvectors can be computed. We present a scalable extension of NEO to pixel-based domains, built on three contributions. First, we use a novelty-weighted continuous Laplacian graph-drawing objective, which enables RL with continuous observations. Second, we embed the resulting eigen-potential options within a hierarchical reinforcement learning framework, enabling coherent temporally extended behavior. Third, we observe that learned eigen-potential rewards are directional but locally unreliable under online approximation; we therefore augment each option reward with a novelty bonus, a novel design idea that proves essential for stabilizing option learning while preserving novelty-directed exploration. Together, these contributions yield stronger and more persistent exploration, enabling longer option rollouts and better access to hard-to-reach novel states. Empirically, our method significantly outperforms both the prior scalable Laplacian-option baseline and a direct extension of NEO on sparse-reward benchmarks under a fixed budget. On Montezuma's Revenge, our best variant achieves approximately 1.8x higher return than both baselines. On Venture, both baselines yield returns near zero, whereas our method achieves a return of 1135. Across seven hard ProcGen games, our method achieves approximately 3.5x and 5.6x higher aggregate normalized return than the two baselines, respectively.

Modern deep reinforcement learning algorithms often store past sensory observations in a buffer for later replay and learning. However, this process is biologically implausible; animals do not store raw sensory observations. Instead, they store their imperfect understanding of the world and the events in it. An alternative to storing all past sensory observations is to learn directly from the observations as they come in and store the seemingly important parts of the incoming data. Eligibility trace algorithms in reinforcement learning (RL) sweep back updates for multiple steps, improving efficiency and providing update rules that do not rely on storing large data buffers. These algorithms were common and effectively used in the linear setting, but have not been widely adopted in the deep RL setting. Many questions remain open to make these algorithms more effective, even as basic as how to incorporate vector step-sizes. We first highlight how adaptive vector step-sizes need to be incorporated into the eligibility trace and then introduce a new vector step-size algorithm that is more amenable to the forward-backward view equivalence needed to derive an update with eligibility traces. We show that our new algorithms improve over other streaming algorithms across Mujoco and MinAtar environments, and are comparable to methods with replay buffers such as PPO.

We investigate the problem of scaling safe reinforcement learning to massively parallel training regimes, where a large number of environments are executed with short rollout horizons. Although this setting improves data throughput and wall-clock efficiency, it creates a mismatch with constrained Markov decision processes, whose safety constraints are defined over full-episode cost returns. We address this mismatch by analyzing staggered environment resets as a phase-mixture distribution over short trajectory segments. This analysis characterizes the bias induced by phase coverage and motivates a phase-aggregated cost estimator that reconstructs episode-level cost estimates from staggered short rollout batches. We further incorporate finite-sample uncertainty and distribution mismatch into a constraint-tightening framework, yielding a scalable Safe RL method compatible with on-policy optimization.Using an MJX-based implementation of Safety Gym, our experiments show that the proposed method achieves performance comparable to non-massively parallel implementations while substantially reducing wall-clock training time

In this paper, we consider mixed-motive games (MMGs) which model various real-world multi-agent scenarios where self-interested agents must both cooperate and compete. We identify an issue with existing gifting methods, where agents become overly dependent on incoming gifts and fail to learn reward-yielding behaviors, a phenomenon we call *gift over-reliance*. To address this, we propose a refusal-augmented gifting framework that balances cooperation and competition in MMGs. Our approach introduces an adaptive refusal mechanism that allows agents to selectively reject a portion of received gifts, thereby mitigating gift over-reliance and preventing the agents from being exploited by other agents. This mechanism is gifting-method-agnostic, serving as a plug-in for existing schemes while enabling decentralized learning. We demonstrate that across diverse MMG environments and gifting schemes, our mechanism consistently reduces gift over-reliance, prevents lock-in to sacrificing policies, and improves $\alpha$-fairness without necessitating major changes to existing training pipelines.

Post-training quantization is a standard tool for the efficient deployment of large language models (LLMs). In safety-aligned models, it introduces a critical and underexplored risk: preserving utility under compression does not ensure that safety alignment is retained. Small perturbations from quantization can disproportionately degrade refusal to harmful prompts, leading to increased compliance with harmful prompts or over-refusal of benign prompts, even when aggregate performance appears stable. This gap reveals a fundamental limitation of existing quantization approaches, which primarily optimize for utility while overlooking failures due to safety alignment degradation. This paper presents ACQueReLlo, a reinforcement learning framework for mixed-precision post-training quantization. It formulates block-wise bit allocation as a constrained Markov decision process and learns a quantization policy under explicit constraints on utility, harmful prompt refusal, benign prompt over-refusal, and overall bit budget. To ensure tractable training, the policy relies on proxy evaluations over sampled subsets of utility, harmful, and benign prompts. Experiments on various LLMs show that the learned policy achieves a better overall balance between compression, utility, and safety alignment than uniform quantization and other baselines.

We propose a primal-dual policy optimization method for reinforcement learning in semi-infinitely constrained Markov decision processes (SICMDP), which extends standard constrained Markov decision processes by allowing a continuum of constraints. We call our proposed method \textbf{S}emi-\textbf{I}nfinitely \textbf{P}rimal-\textbf{D}ual \textbf{P}olicy \textbf{O}ptimization (SI-PDPO). By introducing an infinite-dimensional Lagrange multiplier, we derive the saddle-point formulation of the original SICMDP problem and prove strong duality under mild conditions. We then introduce a regularization term for the dual variable, and solve the resulting regularized saddle-point problem using first-order methods with a policy optimization subroutine (for example, NPG or PPO). Unlike existing primal-type policy optimization algorithms for semi-infinitely constrained reinforcement learning, our method does not require solving an inner-loop optimization problem, thereby mitigating computational intractability and reducing the over-conservative bias. We perform a series of numerical experiments, including a real-world power control task in wireless communications, to evaluate the performance of SI-PDPO. Empirical results show that SI-PDPO consistently outperforms existing primal-type algorithms.

Quantum reinforcement learning (QRL) has emerged as a promising approach for controlling quantum systems and accelerating learning tasks with quantum resources. However, most existing QRL formulations still retain a classical reinforcement learning interface: states and actions are often encoded in computational bases and rewards are extracted through measurement. This is less natural for fully quantum environments in which the relevant state information may live in an unknown basis and intermediate measurements can disturb the trajectory. This motivates a fully QRL framework in which the agent, environment, rewards, and return accumulation are all represented quantum mechanically. In this paper, we take a step toward this goal and introduce a new QRL framework in which state, action, reward, and reward-accumulation registers evolve without intermediate measurement. In this framework, both the policy and the environment are modeled by *conditional quantum channels*: the policy updates the action register conditioned on the state, while the environment updates the state and reward registers conditioned on the action, without measuring or overwriting the corresponding control register. We then propose a quantum policy-gradient algorithm for parameterized conditional quantum channels with a trainable control basis. We also derive finite- and infinite-horizon gradient estimators using a new parameter-shift rule. The algorithm samples a single shifted policy block from a geometric distribution, yielding an unbiased gradient estimator under bounded rewards. Finally, we establish convergence of the resulting stochastic gradient-ascent procedure to stationary points under standard smoothness and step-size assumptions.

Inverse Reinforcement Learning (IRL) is classically solved using Lagrangian optimization, by jointly optimizing a primal-dual problem with gradient descent to obtain the optimal reward function parameters and corresponding optimal policy. While this algorithmic view of IRL has been highly influential, gradient-based reward updates only paint a partial picture, largely ignoring the geometry of the problem in occupancy space. In this paper, we present an occupancy-space view of reward updates in IRL. We leverage the insight that both the optimal solution and its geometry in occupancy space are known, and show that the geometry along the e-geodesic (straight line in occupancy space) toward the optimal occupancy directly affects the gradient of the IRL dual problem in reward parameter space. This perspective naturally suggests preconditioning the dual gradient with the average Fisher information along the occupancy trajectory. We show that this preconditioned gradient is just a convex interpolation of the old reward and a new target reward, yielding efficient natural-gradient-style updates with little additional computational overhead. We then analyse these occupancy-space reward updates from a Mirror Descent (MD) perspective, and show that policy search over the resulting rewards induces approximate entropic MD on the policy. Finally, we conclude with convergence results of such geometry-aware IRL algorithms.

Recent work shows that preference alignment objectives can be interpreted as divergence estimators between aligned (preferred) \& unaligned (less-preferred) distributions, yielding a principled recipe for designing alignment losses. However, this view has so far been limited to preference-based supervision. We extend it to general LLM alignment, including reinforcement learning with verifiable rewards (RLVR), where alignment feedback is given only as scalar rewards. We introduce $f$-Group Relative Policy Optimization ($f$-GRPO), a class of on-policy RL objectives, and $f$-Hybrid Alignment Loss ($f$-HAL), which combines on-policy reward optimization with off-policy preference supervision. We show that these objectives estimate $f$-divergences between reward-aligned \& reward-unaligned distributions induced by above- \& below-average reward responses, and prove expected reward improvement after alignment. Empirically, $f$-GRPO improves over GRPO on math-reasoning RLVR tasks, while hybrid $f$-HAL mitigates reward hacking in on-policy safety alignment when verifiable rewards are unavailable and learned reward models must be used.

Distributional Reinforcement Learning methods aim to learn the entire distribution of returns, yet current algorithmic approaches are limited to a narrow class of distributions. We propose $\varphi$TD, a method to overcome this limitation through computing the loss in the frequency domain. We demonstrate that this further unifies quantile and categorical approaches to distributional RL, as well as enabling the learning of distributions that were previously impossible to capture, such as mixtures of continuous distributions. We empirically study $\varphi$TD across synthetic benchmarks and large-scale deep RL environments, and demonstrate that the flexibility obtained by wider parametric families often leads to improved performance.