Overview

Policy optimization studies how to update a decision-making rule so it earns higher rewards while meeting constraints such as safety, fairness, or limited data. Common approaches include policy gradients, actor-critic methods, trust regions and mirror descent, as well as policy iteration and newer generative-policy methods. Work spans continuous-control benchmarks such as MuJoCo, offline and multi-agent settings, and language-model training for mathematical reasoning, coding, and tool use, where token-, segment-, or trajectory-level rewards guide updates. A central challenge is balancing effective exploration and sample reuse with stable updates, useful credit assignment, and improvements that carry over from training to deployment.

Latest papers

237 papers18 papers in the last four weeks, down 45% on the four weeks before. 0.1% of all new papers.

Oct 1, 2026

Towards Optimal Policy Improvement

Practical Reinforcement Learning (RL) algorithms learn to solve Markov Decision Processes (MDPs) through iterative policy improvement in the presence of approximate evaluation. We study policy improvement from first principles, defining optimal policy improvement as producing the best policy attainable in a single update under specified constraints. We show that optimal improvement restricted to a set of states is equivalent to solving an induced MDP, characterizing planning with an explicit or implicit model as a path towards optimal policy improvement. Because practical methods commonly solve such induced problems through iterative improvement in the form of greedification, we take steps towards optimal greedification under the central practical constraint of approximate evaluation. We formulate greedification under this constraint as probabilistic decision-making under uncertainty and derive a novel operator that is optimal with respect to the resulting objective. Empirically, the operator and its practical gradient-based approximations improve aggregate performance across GumbelAlphaZero, SAC, ReBRAC and Generalized Policy Iteration, in experiments spanning discrete and continuous actions, model-based and model-free, online and offline RL.
Oct 1, 2026

Grounding Large Language Models in DSGE Simulators for Policy Generation and Forecasting

Large language models can produce economic policy responses that sound reasonable, but this does not show that their actions are consistent with economic dynamics. We test this by placing an instruction-tuned language model inside six Snowdrop-backed dynamic stochastic general equilibrium (DSGE) simulators. At each turn, the model observes the economy and a change in economic discourse, selects a bounded policy action, and receives the next simulated state and an economic reward. We implement a common Python interface for repeated rollouts, persistent shocks, state cloning, and rolling-horizon simulation. This setting creates a long-horizon credit-assignment problem. Policy effects may appear several quarters after an action is taken. PPO has a learned value function that can propagate delayed reward to earlier tokens through generalized advantage estimation. GRPO has no learned value function and instead assigns a group-relative advantage from complete rollout returns. It therefore cannot distinguish which earlier turn caused the outcome; if every rollout receives the same return, the normalized advantage is zero. We use PPO as the primary method and GRPO as a matched critic-free baseline. The experiments also test directional semantic signals, reward horizon, trajectory warm starts, cross-simulator transfer, and historically anchored pandemic and monetary-policy shocks. The objective is to judge policy actions by their simulated economic consequences rather than by plausible language alone.
Sep 30, 2026

Role-Adaptive Policy Optimization for Offline Reinforcement Learning

Policy regularization in offline reinforcement learning balances policy improvement against reliance on uncertain value estimates. This balance can differ between selecting actions for execution and supplying actions for critic bootstrapping, yet methods such as TD3+BC couple these roles through a shared policy. We propose Role-Adaptive Policy Optimization (RAPO), which adapts policy-update coefficients according to their roles in value learning and execution. RAPO learns these coefficients by differentiating through candidate policy updates formed using the base algorithm's actor objective. For TD3+BC, RAPO separates bootstrap and execution actors and adapts their coefficients independently: the bootstrap objective penalizes policy-induced changes in target values, while the execution objective evaluates a local policy-improvement surrogate. For IQL, whose value learning is already independent of the execution actor, RAPO preserves the original value updates and adapts only the inverse temperature in advantage-weighted policy extraction. Experiments on D4RL locomotion and AntMaze tasks show improvements over both base algorithms, with larger gains for TD3+BC, whose RAPO instantiation outperforms baselines on average.
Sep 30, 2026

Fast Regularized Policy Mirror Descent with One-Step TD Updates

Policy mirror descent (PMD) enjoys fast convergence in regularized Markov decision processes (MDPs), but existing guarantees often rely on exact or increasingly accurate policy evaluation. We analyze PMD coupled with a persistent critic advanced by one temporal-difference (TD) update. For finite discounted MDPs, we establish global linear convergence in value for exact coordinate-wise Bellman updates, with any positive constant actor stepsize and arbitrary finite critic initialization. The proof combines a resolvent-based auxiliary distribution with a decaying Bellman-violation correction and a potential weighted by inverse coordinate weights. We then study stochastic TD-PMD with general strongly convex mirror maps under a single off-policy Markov trajectory. With suitably chosen constant stepsizes and a finite-batch TD update, the method achieves an expected value gap of εε after O~(1/((1−γ)5σ~bε))\widetilde{O}(1/((1-γ)^5 \widetildeσ_b ε)) transitions. The stochastic analysis relies on the trajectory-wise Lipschitz continuity of the regularizer, derived from uniform bounds on vertex Bregman divergences, together with a visitation-weighted resolvent estimate for signed critic-error propagation that yields an inverse-linear dependence on behavior coverage σ~b\widetildeσ_b. In contrast to many prior guarantees for regularized policy optimization, our sample-complexity guarantee holds without trajectory resets, generative-model access, or nested policy-evaluation loops. Numerical results are consistent with the theoretical convergence analysis.
Sep 30, 2026

Self-Evolving Algorithm-Design Agents: Escaping In-Context Evolutionary Stagnation via Population-Curated Policy Optimization

Large language models are increasingly participating in complex real-world tasks in the form of algorithm-design agents, designing and refining algorithms. Many successful algorithm-design agents adopt pure in-context evolutionary frameworks, but they may quickly plateau in domains that require specialized knowledge. Parametric adaptation offers a way to internalize specialized knowledge, but conventional training requires abundant domain-specific corpora while high-quality algorithms are scarce in complex algorithm-design scenarios. In this paper, we propose sample-efficient parametric self-evolution where agents can explore and learn from self-generated algorithms. First, we characterize in-context evolutionary stagnation and analytically propose the Improvement Chain proposition, showing how learning successive self-generated algorithms can locally increase the likelihood of neighboring algorithms. Motivated by this local-transfer perspective, we further propose Population-Curated Policy Optimization (PCPO) to utilize a global population and a hybrid policy update scheme for retaining and reusing high-quality, diverse self-generated algorithms, shifting the policy towards stronger algorithms. In the task of learning rate schedule design for global placement in electronic design automation, trained only on 4 chip cases, PCPO outperforms the state-of-the-art in-context evolutionary methods (e.g., OpenEvolve and ShinkaEvolve) on average across 16 chip cases. With an 8B-size base model, PCPO achieves competitive performance compared to frontier closed-source models such as GPT-5.5. PCPO also reduces inference-time token cost by internalizing grounded domain knowledge and prompt distillation. Moreover, PCPO achieves significant speedups on four GPU kernel designs, with an average of 8.27×\times speedup against the PyTorch Eager baseline.
Sep 29, 2026

Achieving an O(1/N)O(1/N) Optimality Gap in Average-Reward Weakly-Coupled MDPs

We study average-reward weakly-coupled Markov decision processes (WCMDPs), where a WCMDP consists of NN smaller MDPs, called arms, that share multiple per-step budget constraints. We consider the setting where the arms have identical model parameters, multiple actions, and state- and action-dependent costs. For restless bandits (RBs), a well-studied special case of WCMDPs, prior work has developed policies that achieve an O(1/N)O(1/\sqrt{N}) optimality gap under general conditions, and has further identified conditions under which policies can achieve a better-than-1/N1/\sqrt{N} optimality gap. However, for general WCMDPs, no prior result achieves an optimality gap better than 1/N1/\sqrt{N}. In this paper, we identify conditions analogous to those for RBs under which a better-than-1/N1/\sqrt{N} optimality gap is achievable, and design a policy that attains an O(1/N)O(1/N) optimality gap. Notably, unlike prior approaches based on generalizing priority orderings, our policy is not priority-based but rather is designed to induce locally linear mean-field dynamics.
Sep 29, 2026

Bellman-Certified Rounding for Sparse Policy Deployment in MDPs

Continuous policy optimization may spread an update across many states, even when deployment permits only a few complete state-level changes. We study how much discounted return can be retained when continuous row mixtures are rounded to sparse binary policies in finite MDPs. Policy-dependent visitation couples the row edits, while long horizons make global curvature bounds conservative. From 2d+22d+2 Bellman solves, we derive reusable envelopes that support uniform and candidate-specific guarantees before rounding. A rank-two rational representation of each exchange further permits weighted curvature integration along the realized trajectory. We prove that linear dimension dependence is unavoidable when the budget scales, and that exact global curvature thresholding is hard. Candidate-specific bounds raise pre-rounding certification coverage from 48.2%48.2\% to 74.1%74.1\% on the structured suite. At γ=0.95γ=0.95, local integration lowers the median bound-to-loss ratio from 402.3402.3 to 2.082.08 on coupled instances.
Sep 29, 2026

Markovian Nonconvex ADMM for Reinforcement Learning: Bellman-Resolvent Stability Beyond Smooth Blocks

We identify and study a structural mechanism for Markovian nonconvex ADMM in reinforcement learning. Using finite discounted MDPs as a canonical proving ground, we show that the discounted Bellman resolvent (I−γPπ)−1(I-γP_π)^{-1} can provide the multiplier stability that classical nonconvex ADMM analyses often obtain from a designated smooth block. Starting from this mechanism, we establish convergence under controlled Markov sampling and then under stochastic observations using an empirical Bellman surrogate that jointly represents the random residual and its Jacobian. Markov mixing, initialization drift, observation noise, and decaying bias enter as one operator perturbation, avoiding unbiased product and double sampling requirements. When the perturbations are square summable, the true KKT residual converges almost surely to zero. Under a finite conditional fourth moment condition, a companion iterate satisfies E[G~K+1]≤A/T+(B/T)∑k<Tmk−1,\mathbb{E}[\widetilde G_{K+1}] \le A/T+(B/T)\sum_{k<T}m_k^{-1}, which becomes O(T−1+T/N)O(T^{-1}+T/N) for total Markov sample budget NN, giving O(ε−1)O(ε^{-1}) iteration complexity and O(ε−2)O(ε^{-2}) sample complexity for squared KKT accuracy εε. Beyond stationarity, discounted occupancy coverage yields J⋆−J(π)=O(G)J^\star-J(π)=O(\sqrt G) for direct tabular policies, so covered exact KKT points are globally optimal, while a statewise quadratic Bellman-improvement condition sharpens the relation to O(G)O(G). Finally, nonlinear policy, projected Bellman, and explicit occupancy formulations exhibit the same chain of operator invertibility, dual representation, and multiplier stability. This supports discounted operator invertibility as a reusable structural principle for primal-dual reinforcement learning.
Sep 29, 2026

Towards Better Training Signal: Advantage Clipped Policy Optimization

Reinforcement learning (RL) has become a cornerstone for improving the reasoning capabilities of large language models (LLMs), but the need for on-policy data substantially limits training efficiency. Reusing off-policy data through importance sampling (IS) can improve efficiency but introduce considerable instability. Hence, algorithms such as PPO and GRPO widely adopt IS-ratio clipping to stabilize training. However, training stability and gradient estimate are mainly determined by the product of IS ratio and advantage. To further stabilize training, we propose ACPO, which clips the product of the IS ratio and the advantage, leading to more stable gradient estimates. We also establish a connection between ACPO and gradient clipping in policy mirror descent (PMD), which is a standard technique to stabilize optimization process, and prove the convergence of clipped-PMD under the standard RL setting. Experiments on widely used mathematical reasoning benchmarks show that ACPO consistently outperforms PPO and GRPO in both accuracy and training efficiency, delivering 4-6 percentage points gains on standard math benchmarks, with Qwen3-8B+PPO. Hence, ACPO is a practical and effective alternative to conventional IS-ratio clipping for RL post-training of LLMs.
Sep 29, 2026

Diffusion Policy Improvement with Proposal-Conditioned Refinement Flows

Diffusion and flow policies can model complex behaviors in offline reinforcement learning (RL). However, penalizing their KL divergence from the behavior policy can discourage actions having high critic values with low behavior density. Directly refining behavior proposals may be an alternative, yet Gaussian or deterministic editors limit expressiveness to represent multiple separated modes for the same proposal. In this work, we introduce Proposal-Conditioned Refinement Flows (PReFlow), a policy extraction method combining critic-based proposal selection with a conditional refinement flow. To optimize proposal selection and refinement together, we formulate a KL-regularized objective whose optimum induces a Gibbs policy over final actions under a Gaussian-smoothed behavior prior. The refinement flow can represent multiple high value modes, while a proposal-centered Gaussian reference regulates large action changes. This Gaussian reference further enables us to make use of simulation-free, closed form adjoint matching targets from sampled endpoints and critic gradients, yielding a single velocity regression loss without a backward adjoint solve. On 50 OGBench tasks, PReFlow achieves competitive offline performance and the highest aggregate score among the compared methods after online fine-tuning, reaching 91% after 500K environment steps.
Sep 29, 2026

Regularized policy gradient with learned mixtures of Gaussians for games with continuous actions

Most successes of superhuman game-playing algorithms are in games with discrete actions, yet in auctions, robotics, sports, or trading, actions are nearly continuous. Prior techniques either rely on expert-designed discretizations or are sample inefficient. We present a scalable policy-gradient algorithm for large sequential games with continuous or mixed discrete and continuous actions. It combines magnetic mirror descent with a mixture of Gaussians reparametrization, trained via self-play. We show that it approximates equilibrium in games where gradient descent fails. In sequential games, it outperforms neural fictitious self-play and matches or outperforms the final strategies of policy space response oracles with 3.5--5.5×\times fewer samples. In heads-up no-limit Texas hold'em, it performs on par with Slumbot.
Sep 29, 2026

Factorized Scheduling Principle: Learning Interpretable and Transferable Policies via Structured Additive Functions

Scheduling problems arise from repeatedly selecting one item from a set of candidates based on their states. These problems often reduce to assigning priority scores and choosing the highest-ranked item. In this work, we propose a factorized scheduling principle (FSP) framework to learn interpretable and transferable scheduling rules. The FSP framework represents system states as condition distributions and decomposes a global scheduling principle into additive univariate and pairwise components with identifiability constraints. The scheduling principle enables the framework to maintain a simple priority-based structure during deployment. This principle is learned by using a policy-based objective combined with a temporal-difference signal defined on the condition distribution. Experiments on synthetic and realistic scheduling tasks demonstrate the FSP framework's strong performance, interpretability, and zero-shot generalization across different system scales.
Sep 28, 2026

RareDx: Controlled Knowledge Integration and Graph-Grounded Policy Optimization for Rare-Disease Diagnosis

Rare-disease diagnosis is a long-tail reasoning problem: phenotypes are incomplete, individual disorders are sparsely documented, and relevant evidence is distributed across ontologies, gene annotations, and biomedical text. Language models consequently favor common conditions, miss rare candidates, or produce plausible but invalid names. We introduce RareDx, which couples controlled evidence use with knowledge-graph-grounded policy optimization. RareDx-Harness normalizes heterogeneous records into one ranked-diagnosis task and compares direct inference, static retrieval, adaptive tools, and structured phenotype-gene-disease reasoning over a shared knowledge layer. The training pipeline combines Top-10 post-training with RareDx-KGPO, our knowledge-graph-grounded policy optimization method. Its reward projects predictions into a canonical disease graph and integrates curated graded relevance, ontology proximity, biomedical similarity, and phenotype consistency. Vocabulary and output-budget constraints prevent dense partial credit from rewarding fabricated or overlong differentials. Across eight benchmarks, the complete RareDx system centered on Qwen3.5-9B reaches 38.34 macro Hit@10, 1.60 points above GPT-5.5 under the archived protocol; a disjoint validation-selection audit retains a 6.80-point routing gain over Direct on held-out cases. The 27B system reaches 23.53/36.56/40.76 at Hit@1/5/10. Controlled ablations show that retrieval is not uniformly helpful and that controlled routing is central to the gain. These results indicate that structured medical knowledge can turn a compact model into a competitive diagnostic ranker across heterogeneous long-tail settings in clinical practice.
Sep 28, 2026

ReSPO: Reshaped Sequence Policy Optimization for Gradient Starvation in Off-Policy Learning

Reinforcement learning from verifiable rewards (RLVR) frequently reuses rollouts across multiple policy updates, increasing the mismatch between the current policy and the data-generating policy. We identify a sign-dependent gradient starvation problem in clipped policy optimization: clipping suppresses under-generated positive responses at the low-importance-weight tail while permitting severely over-generated negative responses to dominate the high-weight tail. To address this, we propose ReSPO (Reshaped Sequence Policy Optimization), which replaces clipping with a smooth, two-branch sequence-level kernel derived from an αα-divergence variational objective and an exponential variance-control tilt. The positive branch preserves a nonzero gradient weight for under-generated positive responses, while the negative branch suppresses heavily over-generated negative responses. We demonstrate that ReSPO effectively learns from long positive reasoning trajectories during early training, even when accumulated policy drift relegates them to the low-importance-weight tail. On dense and MoE Qwen3 models, ReSPO accelerates early optimization, improves final training scores, and achieves higher held-out benchmark performance under a rollout reuse, validating our approach on importance-weight tail control in off-policy learning.
Sep 28, 2026

Proactive Dialogue Policy Optimization via Cognitive-State Transition

Proactive dialogue requires agents to continually adapt their policies to user feedback while progressing toward task objectives over multiple turns. To move beyond imitation learning on static datasets, recent approaches use user simulators to collect interactive data for policy optimization. However, many simulators do not explicitly model the evolution of user cognition, limiting the consistency and state dependence of feedback across turns. Moreover, representing each action only by a high-level strategy label overlooks the large utterance space and cannot distinguish alternative realizations of the same strategy. To this end, we jointly design a Cog\textbf{Cog}nitive User Sim\textbf{Sim}ulator (Cog-Sim)\textbf{(Cog-Sim)} and C\textbf{C}ognitive-S\textbf{S}tate T\textbf{T}ransition--Driven P\textbf{P}olicy O\textbf{O}ptimization (CSTPO)\textbf{(CSTPO)}. Cog-Sim maintains the user's cognitive and affective states and generates responses through constrained state transitions across turns, so feedback depends on both the realized utterance and the user's current state. CSTPO organizes each action as a hierarchical strategy--utterance representation: a high-level strategy label constrains utterance sampling, and utterances are optimized within each label. Sparse complete-branch sampling reuses shared dialogue prefixes and estimates separate strategy-level and utterance-level advantages, enabling fine-grained optimization at both levels. Across three tasks, Cog-Sim exhibits monotonic dose--response relationships and is preferred over prompt-based simulators for naturalness. CSTPO improves Qwen3-14B's performance to a level comparable to that of GPT-5.5-based planning methods.
Sep 27, 2026

Vanilla Policy Optimization Is Both Optimal and Differentially Private for Stochastic Contextual Bandits

Can vanilla policy optimization explore enough to achieve near-optimal regret in stochastic contextual bandits? We show that standard exponential policy updates driven by offline regression do so under realizability, without exploration bonuses or importance weighting. For AA actions, TT rounds, and a finite prediction class FF, vanilla PO achieves O~(ATlog⁡(∣F∣))\widetilde O(\sqrt{AT\log(|F|)}) regret with high probability. Our analysis reveals an implicit exploration mechanism of independent interest: gradual policy updates prevent actions from losing probability too quickly, allowing the regression oracle to learn their expected losses. We further develop a batched version using only O(log⁡T)O(\log T) regression calls and policy switches, and show how private regression oracles yield differentially private contextual bandit algorithms without composition across batches. For a finite class, this gives pure εpriv\varepsilon_{\rm priv}-DP and regret O~(ATlog⁡(∣F∣/δ)(1+εpriv−1/2))\widetilde O\left( \sqrt{AT \log(|F|/δ)}(1+\varepsilon_{\rm priv}^{-1/2}) \right). Finally, experiments across oracle-based contextual bandit algorithms, with and without privacy, demonstrate the practical effectiveness of policy optimization and the value of explicit exploration under stronger privacy constraints.
Sep 27, 2026

NLPG: Natural-Language Policy Gradients for Self-Evolving Language Agents

Large language model agents increasingly rely on compound programs for retrieval, tool use, reasoning, and verification, yet their failures often arise from local procedural decisions. Existing reinforcement-learning and prompt-optimization approaches typically rely on scalar rewards or repeatedly modify entire prompts, making it difficult to capture and reuse procedural improvements while preserving a frozen agent. To address this problem, We propose Natural-Language Policy Gradients (NLPG), an external policy-memory method for improving a fixed agent without changing its model parameters or program structure. NLPG diagnoses execution traces, propagates downstream feedback backward through the module graph, and converts recurring failures into route-local natural-language corrections that are aggregated into bounded policy updates for subsequent executions. Across six benchmarks covering memory, reasoning, instruction following, and evidence verification, NLPG also outperforms the strongest listed baseline for each benchmark by 8.71 percentage points on average. These results provide evidence that evaluated procedural experience can be transformed into local and interpretable policy updates, enabling continual improvement of frozen agents.
Sep 26, 2026

Learnable Randomization as Commitment Against Adaptive Optimizers

A pricing page can walk the posted price up to the last amount a buyer still accepts, a recommender can hold back a better item for a barely acceptable promoted one, and a classifier can shift its boundary once applicants change their features. The system predicts the response and then picks the menu that serves its own objective, so the surplus above the user's cutoff is taken. Playing the single best action publishes that cutoff, while noise on actions the user would never take throws away payoff and teaches the platform that a worse menu is still acceptable. We study unpredictable near-optimal policies (UNOP), which mix uniformly on near-best actions that remain individually rational. The mixture is a commitment about the response. On a finite price grid, when the best sure-demand price strictly out-earns the randomized band, a seller who already knows the curve posts below the band, and the purchase that occurs is deterministic. Knowing that curve is not the same as predicting the next draw. The mixture can be learned and the optimizer can match its best response, while the user's payoff stays higher because the mixture changes which action is targeted. In pricing and in policy-aware recommendation this leaves more surplus than greedy play when the platform optimizes against the curve and more than one action is acceptable. The gain goes away under quality ranking, a singleton near-optimal set, a wrong utility estimate, or a short-horizon explorer. That is also where mixing should be turned off if the other side is trying to cooperate.
Sep 26, 2026

Action Shaping: Policies Absorb What They Can Express

Reward shaping has a theorem: a potential-based term can be removed without changing the optimal policy. The same practice on the action channel, an offset added in training and dropped at deployment, has no theorem. Nothing cancels an action offset, so the correction is kept at deployment or removed without a guarantee. We call it action shaping and state its principle. A trainable policy absorbs an offset its own output layer can reproduce exactly, which is what we mean by express; what is absorbed can be removed with the return intact. Its minimal instance is a zero-initialized linear head behind a learnable gate, added to an actor that trains through a learned action-value function, with no penalty or schedule. The gate rises and then falls on its own, for deterministic and stochastic actors alike, and on 20 tasks removing the head costs almost nothing. The condition is exact reproduction, not capacity: a nonlinear head with more parameters is not absorbed, and in a paired control, one linear path added to a nonlinear base head restores absorption. Exact reproduction gives the loss a flat direction that gradient noise drifts along, and the offset's amplitude indicates, before removal, what dropping the head will cost. Action shaping thus gains the counterpart of the shaping theorem, a condition for absorption, together with the mechanism behind it and a diagnostic that reads it. Policies absorb what they can express, and only that.
Sep 25, 2026

G2MAF: Test-Time Gradient Guidance for Multi-Agent Flow Policies

Offline multi-agent reinforcement learning (MARL) learns cooperative policies from fixed datasets without further environment interaction and a learned policy is frozen at deployment. Such a frozen policy typically proposes a single joint action and executes it directly at deployment time. However, this one-shot deployment often commits to a suboptimal proposal, even when better nearby alternatives remain consistent with the behavior data. To address this issue, we propose Gradient Guided Multi Agent Flow (G2MAF), a refinement framework for optimizing joint policies at test-time. G2MAF applies one globally normalized, projected critic gradient to guide and coordinate all agents' corrections while keeping the action both feasible and close to the frozen policy proposal. Across 24 MPE and SMAC settings, its canonical variant improves 20 frozen settings, with mean relative gains of 9.2% on MPE and 8.9% on SMAC, with model inference latency increased by about 6% only.
Sep 24, 2026

PolicyAttention: Softmax Attention Implements Policy Mirror Descent for Closed-Loop Control

Can causal softmax attention implement policy mirror descent as a repeated controller rather than a one-step algebraic identity? Negative-entropy policy mirror descent (PMD) has the statewise update PMD⁡η(π,Q)=softmax⁡(log⁡π+ηQ)\operatorname{PMD}_η(π,Q)=\operatorname{softmax}(\logπ+ηQ). Building on the known Q-TD-PMD recursion, we construct one fixed causal-softmax actor--environment--one-step-critic protocol with explicit actor, routing, sampling, and normalization residuals, and propagate them to the policy actually returned. The construction states the finite-logit/full-support domain, the external tokenization and sampling boundary, and the mean-zero LayerNorm carrier conditions required by the normalized compilation. Separately trained pre-LN Transformers recover the target computation empirically. A frozen one-step audit model is closest to PMD among the tested fixed rules; in a preregistered five-run S=4S=4 repeated-control test, the learned actor with an exact one-step critic reaches median returned-policy loss 1.052×1.052\times the Exact PMD oracle and retains the criterion across four no-retraining shifts. The same checkpoints with their learned critic give descriptive median 1.050×1.050\times the oracle (no registered margin). At S=8S=8, replacing the exact critic by the learned critic raises median T=20T=20 loss to 0.02250.0225 yet leaves the Liang--Lai and Algorithm Distillation adaptations 20.220.2--24.2×24.2\times higher-loss; this is a one-sided sampled-critic bound because PolicyAttention consumes 144 generative transitions per round versus 20 on-policy transitions for the adaptations. The strict 20-transition comparison remains open. At S=8,16S=8,16, the exact-critic common-harness comparison remains 17.717.7--28.2×28.2\times lower-loss than those adaptations, with the information asymmetry stated locally.
Sep 19, 2026

Learning When to Refine: Long-Horizon Reinforcement Learning for Budgeted Neural-Operator PDE Solvers

Neural operators provide fast surrogates for time-dependent PDEs, but autoregressive deployment creates a refinement-allocation problem: prediction errors vary over space and time, while only a finite number of local corrections can be committed along a trajectory. We formulate this as budgeted adaptive neural-operator solving. A global Fourier neural operator advances the full field, a local operator proposes patch-wise residual corrections, and a set-aware selector chooses where to refine. A macro policy decides when and how much of the remaining refinement budget to spend. We introduce rollout-verified policy improvement (RV-PI), which evaluates feasible refinement counts through actual continuation rollouts of the learned PDE solver, converts long-horizon advantages into conservative policy targets, and accepts an update only when held-out trajectory error improves. On the shallow-water benchmark with a 32-intervention budget, RV-PI achieves a three-seed mean trajectory relative L2 error of 0.6910, improving over immediate-only policy improvement by 5.37% and RandomMacro by 2.41%. On the forcing-driven Brusselator benchmark with a 76-intervention budget, RV-PI attains 0.09954, improving over immediate-only policy improvement by 2.31% and RandomMacro by 5.32%. These results show that, under a fixed refinement budget, the value of a local correction depends on its downstream effect on the autoregressive trajectory, not only on its immediate error reduction.
Sep 18, 2026

GVPO++: Group Variance Policy Optimization for LLM Post-Training and On-Policy Distillation

Post-training plays a pivotal role in enhancing the reasoning capabilities and task-specific expertise of large language models (LLMs). Despite recent advances in post-training methods, such as Group Relative Policy Optimization (GRPO), their practical deployment remains impeded by training instability arising from the reliance on importance sampling. We introduce Group Variance Policy Optimization (GVPO), a novel post-training method that integrates the analytical solution of KL-constrained reward maximization into its gradient weighting scheme. This formulation provides an intuitive interpretation: GVPO's gradient corresponds to the mean squared error between the central distance of implicit rewards and that of actual rewards. GVPO offers two key advantages: (1) it guarantees a unique optimal solution, exactly to the KL-constrained reward maximization objective, and (2) it enables flexible sampling distributions without requiring importance sampling. Beyond general post-training, we show that GVPO naturally extends to on-policy distillation (OPD). Furthermore, GVPO enables the optimization of a broad family of extended OPD objectives, providing a principled foundation for diverse objective design. By unifying theoretical guarantees with practical adaptability, GVPO establishes a new paradigm for reliable and versatile LLM post-training and on-policy distillation.
Sep 17, 2026

Accelerating Visual Policy Learning with Sampling-Based Model Predictive Control

Learning visual policies for locomotion and manipulation requires coordinating contact with the environment and can incur substantial computation and GPU memory costs. First-order policy gradients (FoPG) reduce training cost through differentiable simulation, but local optimization can converge to unintended contact patterns. To address this shortfall, we propose Sampling-Guided Policy Search (SGPS), which couples recurring action-target refinement by sampling-based model-predictive control with first-order policy optimization. Behavior cloning initializes the policy from sampled actions; training then alternates sampling-based refinement with short-horizon FoPG updates under perturbed initial states and randomized dynamics. For visual policy training, we use a decoupled FoPG formulation that excludes rendering from the computation graph, enabling direct learning from depth observations without a state-policy teacher. On a single GPU, SGPS learns policies for locomotion, obstacle traversal, crate pushing, and bimanual carrying on simulated Unitree Go2 and G1 robots. Our experiments further show that refinement improves policy learning beyond initialization and tracking alone. For hardware deployment, the distilled policy transfers zero-shot to a real Go2 and uses onboard depth to autonomously trot, crawl, clear hurdles, and switch between these behaviors.
Sep 17, 2026

Dual-Axis Policy Optimization for LLM Agents: Bayesian Feedback Attribution and Trajectory Mass Normalization

Reinforcement learning for LLM agents involves two distinct optimization di- mensions: how environment feedback is exploited within a trajectory, and how complete trajectories are aggregated across a batch. We formulate these dimen- sions as Intra-Trajectory Feedback Attribution and Inter-Trajectory Objec- tive Aggregation, and introduce BATON (Bayesian Attribution and Trajectory Objective Normalization), a dual-axis policy optimization framework. BATON instantiates the first axis with Bayesian Feedback Attribution, which constructs a feedback-conditioned posterior over sampled actions, and the second with Trajec- tory Mass Normalization (TMN), which assigns equal optimization mass to com- plete trajectories. Experiments with GRPO and GiGPO on ALFWorld, WebShop, and SearchQA show that both axes provide independent gains and that their combi- nation consistently achieves the strongest overall performance across model scales.
Sep 14, 2026

Bellman Policy Optimization

Reinforcement learning with verifiable rewards (RLVR) improves the reasoning capabilities of large language models (LLMs). We introduce Bellman Policy Optimization (BPO), a critic-free method derived from Policy Mirror Descent (PMD). For autoregressive generation with terminal rewards, BPO uses the Bellman equations to reformulate PMD as a trajectory-level objective. The reformulation avoids estimating state values at intermediate states. We prove that it has the same unique optimal solution as the original PMD objective. We derive the practical BPO loss by approximating this objective. Its mismatch-correction weight is a smoothed ratio of complementary token probabilities. Experiments on mathematical reasoning benchmarks demonstrate the effectiveness of BPO.
Sep 14, 2026

HISPO: Hierarchical Importance-Sampling Policy Optimization with Entropy-Derived Segments

Reinforcement learning with verifiable rewards (RLVR) has become a central approach for improving mathematical reasoning in language models, but long-form completions introduce a difficult credit-assignment problem: different parts of a solution trace may contribute unevenly to final correctness. Existing policyoptimization objectives for RLVR commonly apply importance-sampling correction at either the token level (GRPO, DAPO) or the sequence level (GSPO), imposing different granularities for assigning credit across a response. We introduce Hierarchical Importance-Sampling Policy Optimization (HISPO), a segment-level policy-optimization method that constructs rollout-time entropy-derived contiguous segments, assigns soft entropy-based saliency weights, and applies clipped importance-sampling correction at the segment granularity. This provides an intermediate correction unit between token-level GRPO/DAPO and sequence-level GSPO. We evaluate HISPO by fine-tuning Qwen3-1.7B-Base on mathematical reasoning tasks. Across six benchmarks, HISPO improves Pass@8 over the strongest baseline on all benchmarks and matches or exceeds the strongest baseline in Acc@8 on five of them. On AIME25, HISPO improves over GRPO by +3.75 Acc@8 and +3.78 Pass@8, and over GSPO by +2.50 Acc@8 and +1.27 Pass@8. These results suggest that segment-level correction is a promising granularity for RLVR in long-form mathematical reasoning.
Sep 14, 2026

Refinement-Based Flow Policy Optimization

Flow-based policies offer an expressive representation for online reinforcement learning, but conventional flow matching requires samples drawn from the distribution to be modeled. This poses a challenge when the desired action distribution is defined only implicitly by a Q-function, since directly sampling actions from the resulting distribution is generally intractable. We propose Refinement-Based Flow Policy Optimization (RFPO), a novel framework for training a flow policy in online reinforcement learning by alternating between Q-guided sample refinement and self-target flow matching. RFPO first generates actions from Gaussian noise using the current flow policy and then uses a finite-step stochastic refinement procedure to move them toward an energy-based distribution induced by the Q-function. Each refined action is then paired with its corresponding initial noise sample and used as a fixed target for flow-matching training. By repeatedly refining its own outputs and learning from the resulting targets, RFPO incorporates Q-guidance into the policy without requiring direct samples from the target distribution, while retaining the capacity to represent multiple action modes. We further provide a theoretical analysis of the distributional dynamics induced by RFPO. Across six continuous-control tasks, RFPO matches or outperforms a standard Gaussian-policy baseline on almost every task. Experiments on six synthetic two-dimensional target distributions with diverse geometries demonstrate that RFPO captures complex multimodal structure without mode collapse.
Sep 9, 2026

Why Sample What You Can Enumerate? Exact Policy Optimization for Genomic Tool Selection

Reinforcement learning over a frozen reasoner has become a common recipe for teaching a policy which external tools to invoke. We show that this recipe becomes structurally mismatched in specialist scientific settings where the complete tool-subset space is enumerable. There, a small set of recurring computational capabilities covers the domain, so the space of tool subsets is combinatorial yet small enough to enumerate, and GRPO still estimates an action expectation from a handful of sampled rollouts. Worse, the approximation degrades as training succeeds: as the policy concentrates on preferred subsets it resamples them, sampled rewards collide, and the group-normalized advantage vanishes. On genomic reasoning the fraction of questions yielding no reward signal rises from 0.2% under a uniform reference policy to 20.8% after GRPO training. As a remedy, we introduce FGPO (Full-Group Policy Optimization), which (1) scores every tool subset and optimizes the exact action expectation, so each update sees the complete action space, and (2) precomputes the reward of each question--subset pair into an exhaustive table, removing frozen-reasoner calls from the training loop entirely. Across five frozen reasoners and three genomic benchmarks, FGPO outperforms GRPO in all 15 settings by 6.75 points on average and up to 14.20, while a standard on-demand GRPO schedule would require 2.4 times as many frozen-reasoner reward evaluations and, on GenomeQA, FGPO cuts invoked tools per question from 2.36 to 1.40.
Sep 8, 2026

Entropy-Regularized Rank-Masked Policy Optimization for Test-Time Reinforcement Learning in Code Generation

Existing methods for test-time reinforcement learning (TTRL) derive rewards from answer-level self-voting on unlabeled test-time tasks with canonical answers, but this breaks down for code generation because programs cannot be compared by surface form and therefore do not directly provide a usable training signal. To make TTRL applicable to code generation, we propose probe-driven TTRL, which constructs output-free probe inputs from the problem statement, executes candidate programs on these probes, and defines a Probe Consensus Reward (PCR) from the resulting behavioral agreement. PCR provides a behavioral training signal for open-vocabulary programs, but it is not a fully reliable verifier and remains susceptible to reward hacking through spurious consensus. We therefore introduce Entropy-Regularized Rank-Masked Policy Optimization (ERPO), which converts low PCR into conservative negative updates through rank masking and controls policy drift with an entropy ceiling. On coding benchmarks, ERPO substantially improves pass@1 and pass@k in both in-domain adaptation and zero-shot transfer.