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 51% on the four weeks before. 0.1% of all new papers.

Nov 12, 2026

Improved Regret via Variance-Aware Policy Optimization in Adversarial MDPs

We study episodic adversarial MDPs with unknown dynamics and full-information rewards and achieve regret with leading term O~(H3SAK)\widetilde O(\sqrt{H^3SAK}). Here HH is the horizon, SS the total number of states across decision layers, AA the number of actions, and KK the number of episodes. This improves the two incomparable prior leading terms by factors S/H\sqrt{S/H} and H3\sqrt{H^3}, leaving a factor H\sqrt H to the lower-bound scale. We give a reduction from unknown to known dynamics for downstream-dependent algorithms, whose policy at step hh is computed only from information at steps h+1h+1 and onward. This reduces the remaining leading-order gap to finding a downstream-dependent known-dynamics algorithm with O~(HK)\widetilde O(H\sqrt K) regret, which would yield the tight O~(HSAK)\widetilde O(H\sqrt{SAK}) leading term for unknown dynamics. We instantiate the reduction with VOPO, a policy optimization (PO) algorithm. Its improved guarantees follow from a variance-aware analysis enabled by a novel bonus, which we call the Stability-Matching Bonus (SMB). VOPO achieves O~(H3K)\widetilde O(\sqrt{H^3K}) regret under known dynamics with full information; its H\sqrt H gap to the target HKH\sqrt K rate is inherited by our unknown-dynamics guarantee. We also show that with known dynamics and trajectory bandit feedback, VOPO achieves O~(H2SAK)\widetilde O(\sqrt{H^2SAK}) regret. Both known-dynamics guarantees improve the corresponding state-of-the-art PO bounds by a factor H\sqrt H. A matching lower bound establishes that the full-information rate is tight in one parameter regime.
Nov 10, 2026

A Theory of Policy Optimization in Post-Training: Coverage Collapse and Reference Resets

Policy optimization is central to language model post-training, yet principled guidance remains limited. We revisit its theoretical foundations to identify algorithms that learn near-optimal policies using total samples and gradient steps governed by the base model's response coverage, a measure of how readily it generates optimal responses. We call such algorithms coverage-preserving. This guarantee captures the goal of amplifying capabilities accessible through sampling from the base model. (1) Coverage collapse. REINFORCE, the basis for PPO and GRPO, can drive the probability of optimal responses to near zero, even as reward improves. This coverage collapse stalls optimization, and convergence can require exponentially many steps despite good initial coverage. The failure arises even with exact gradients and persists under KL regularization to the base model. (2) Reference resets. We show that repeatedly resetting the KL reference to the most recent policy mitigates coverage collapse. We formalize this intervention in Approximate Policy Mirror Descent (APMD), a meta-algorithm that approximately solves a sequence of KL-regularized RL subproblems. APMD enables stronger regularization, making each subproblem easier to solve, while reference resets allow convergence to global optimality. (3) Coverage-preserving algorithms. For softmax linear policies over pretrained features, we use APMD to establish the first global convergence guarantees showing that widely used families of post-training algorithms are coverage-preserving. These include stylized PPO and GRPO variants with explicit KL regularization and reference resets, and squared-loss algorithms that implement them implicitly, including those used to train the Kimi models. These results clarify how regularization, clipping, and objective choice affect the sampling and optimization costs of post-training.
Nov 9, 2026

Audited Gradients, Anchored Policies: Model-Based Reinforcement Learning for Efficient Flow Control

Feedback controllers for flows are expensive to design, since each candidate must be evaluated with a high-fidelity solver. Differentiable surrogates reduce that cost, but they also introduce a failure mode that prediction error cannot detect. Policy optimization uses the derivative of the objective with respect to the action, which we call the action gradient and which can be computed either through the surrogate or through the high-fidelity solver. Optimization through the surrogate uses only its action gradient, and a surrogate can follow the flow trajectory closely while misestimating this derivative. The surrogate action gradient is compared against a central finite-difference estimate from the high-fidelity solver on held-out states, under acceptance criteria fixed before the comparison, and we call that comparison an audit. A low-dimensional feedback law is then fitted to the surrogate action gradient when the surrogate passes the audit, and to the solver estimate when it does not. This law initializes a neural policy, and a quadratic penalty holds the policy near it during optimization. The proposed approach is validated on four test cases. The resulting controllers outperform or match a model-free baseline, while using 3 to 156 times fewer high-fidelity solver interactions across the four test cases. One surrogate predicts the objective at correlation 0.9996 but its action gradient at only 0.459, so prediction accuracy does not identify a usable surrogate. Another surrogate, whose action gradient agrees with the solver estimate at correlation 0.996, makes the objective 0.52% worse when the policy is optimized through it without the anchor penalty, so an accurate gradient alone does not provide a working controller. Finally, a surrogate that passes the audit at one fixed action fails it around the actions the deployed controller applies, so a pass is local.
Nov 6, 2026

Ordered Policy Optimization

Proximal Policy Optimization (PPO) stabilizes policy updates by clipping the importance ratio between the new policy and the data-collecting policy. The standard rule uses one fixed clipping range for all samples. While this choice offers simplicity and effectiveness, it fails to distinguish statistically harmful extreme ratio values from large, yet tolerable changes. We study how to adapt the clipping radius to the reliability of the induced importance ratio distribution. Our case study demonstrates that reliable updates require controlling these extreme ratios, rather than merely verifying that individual ratios fall within a fixed interval. Motivated by this observation, we propose Ordered Policy Optimization (OPO). In each minibatch, OPO sorts the importance ratios, selects the extreme ratios that potentially have the largest effect on the update, and assigns them adaptive sample-wise radii instead of using the fixed clipping radius. These radii are derived from a prescribed tail profile, which specifies desired target values for the selected extreme ratios. The resulting objective remains first-order and requires only minibatch sorting and a modified policy loss. Experiments conducted on MuJoCo and Atari benchmarks demonstrate that this simple modification achieves strong performance relative to established baselines.
Nov 5, 2026

Foresee to Trade: Forecast-Augmented Policy Optimization for Portfolio Management

Online investment services must translate asset opportunities into investment outcomes for users' accounts, yet the same market opportunity can warrant different account actions. We study how supervision from future asset opportunities can strengthen account-level trading decisions and propose Forecast-Augmented Relative Policy Optimization (FRPO). FRPO separately supervises asset opportunities and account decisions through a jointly learned representation: a Financial State Head (FSH) learns future relative-return targets to provide asset-level opportunity supervision, while Same-Path Group-Relative Optimization compares trading trajectories under the same market path and initial account to provide account-level relative-outcome supervision. An Adaptive Gate coordinates auxiliary supervision using predictive variance and gradient alignment. FSH is used only during training; inference uses the encoder and policy for trading. Across DJIA30, TOPIX Core30, and three investment-style portfolios, ten-seed evaluations show that FRPO achieves the highest mean cumulative and annualized returns among the compared methods in all five settings. These results support strengthening investment agents' trading decisions through joint supervision from asset opportunities and trading outcomes.
Nov 5, 2026

Learning Multi-Agent Orchestration:Higher-Order Interactions in Policy Improvement

Learned orchestration has emerged as a promising paradigm for large language model (LLM)-based multi-agent systems, in which a central orchestrator selects agents and controls the information they receive throughout a task. Despite rapid empirical progress, existing methods train the orchestrator by crediting each decision with its own advantage, implicitly assuming that the other decisions in the episode stay fixed. This assumption fails in orchestration, where the contribution of each decision to the final outcome depends on the decisions preceding it, so the gain of a joint update is not the sum of its per-decision effects. To address this challenge, we introduce IAPO, an interaction-aware policy optimization framework that explicitly accounts for interactions among decisions. Specifically, IAPO exactly decomposes the performance difference between two policies into the effects of individual decisions and their interactions, which we prove can account for half of the attainable gain, and then optimizes both from sampled trajectories while certifying that every update is an improvement. On long-input versions of GSM8K, MATH, DROP, MBPP and HumanEval, against fifteen published multi-agent and workflow-search systems on the same 7B executor, IAPO's deployed plans reach the strongest system's accuracy with fewer tokens, and removing the interaction term costs 7 to 17 points of accuracy on code.
Nov 2, 2026

MAPO: Mixture of Experts with Alternating Policy Optimization for Universal Humanoid Motion Tracking

Universal humanoid motion tracking is essential for enabling humanoid robots to perform diverse tasks in human environments. To capture the wide diversity of human motions, mixture-of-experts (MoE) policies have been widely adopted, leveraging dynamic routing to enhance policy expressiveness. However, we identify that divergence between old and updated policies can be misestimated during policy-gradient updates of MoE policies, due to their dynamic routing. Because most policy-gradient algorithms regulate policy updates based on the divergence, this misestimation can lead to either overly conservative or excessively large updates, which can yield a suboptimal policy. In this paper, we propose Mixture of Experts with Alternating Policy Optimization (MAPO), which alternately optimizes the router and the experts to avoid the divergence misestimation. MAPO consistently tracks 182 motions that PPO with MoE fails to track, whereas PPO with MoE tracks only one motion that MAPO fails to track. MAPO also attains the highest tracking success rate among the compared methods and the lowest tracking error across most error metrics, demonstrating its superior performance. Videos can be found on the project website: https://mapo-corl2026.github.io/MAPO/
Nov 2, 2026

Do Visual RL Gains Scale at Test Time? A Quantile View of Post-Training

Reinforcement learning (RL) post-training improves the single-sample quality of visual generative models, but deployment often provides additional inference compute through Best-of-N sampling. Whether RL gains transfer under such test-time scaling, particularly on difficult held-out tasks, remains poorly understood in the visual domain. Our central contribution is to systematically identify and characterize a failure mode of scaling transfer in visual RL: across multiple RL methods, model backbones, and evaluations spanning compositional generation, text rendering, and preference alignment, higher single-sample reward does not necessarily imply a better quality-compute frontier. RL gains may attenuate, saturate, or even reverse at large sampling budgets, with the base model eventually outperforming its post-trained counterpart. We explain these regimes through a quantile-transfer framework, showing that Best-of-N places increasing weight on the upper tail of the reward distribution as N grows. Consequently, RL can improve typical samples while compressing or degrading the rare high-quality outcomes that test-time scaling exploits. As a simple intervention motivated by this analysis, we introduce Quantile Utility Policy Optimization (QUPO), which optimizes a controllable mixture of quantile utilities corresponding to different inference budgets. Experiments on SD3.5-M and BAGEL show that QUPO better preserves upper-tail quality and substantially improves scaling on challenging held-out prompts relative to standard RL post-training. Our results suggest that visual generative models should be evaluated and optimized not only for single-sample quality, but for how post-training reshapes the full reward distribution that inference-time scaling exploits.
Oct 31, 2026

Zeroth-Order Policy Optimization for Scheduling in Queueing Networks

Reinforcement learning in queueing network scheduling presents significant challenges because of the countably infinite state space and the high stochasticity in transitions. In this paper, we propose zeroth-order policy optimization algorithms that are easy to implement and rely solely on estimates of the average cost, a single scalar quantity, without requiring the estimation of value or advantage functions. Our analysis focuses on back-pressure guarded policies and establishes uniform smoothness of the average cost objective with respect to the policy parameters. Building on this property, we establish finite-time convergence guarantees to an εε-stationary policy for the proposed algorithms. Numerical experiments on multiple queueing networks, including systems beyond classical exponential service distribution, demonstrate competitive performance, outperforming gradient-based RL methods like PPO and traditional queueing baselines.
Oct 31, 2026

Hi-CoLA: High Confidence Lower Bound Approximation Based Reinforcement Learning for Flex-Route Transit Operation Control

This paper studies safe policy improvement for real-time operation control of flex-route transit (FRT) systems, where robustness is critical and any performance decrease is undesirable. While reinforcement learning (RL) has shown strong empirical success, its adoption remains limited in safety and cost critical applications, such as intelligent transportation, since learned policies lack statistically certified performance guarantees before deployment. In this paper, we propose the High Confidence Lower Bound Approximation (Hi-CoLA) RL framework, which combines importance sampling and concentration inequalities with a differentiable surrogate network to efficiently estimate the high-confidence performance lower bound and optimize it via policy gradient for robust performance improvement. On the FRT operation control problem, Hi-CoLA outperforms state-of-the-art offline and safe RL methods in both average and tail-risk performance. A case study on an actual route in Chicago, implemented in the SUMO simulator, serves as a proof of concept. Hi-CoLA enhances real-time FRT control with guaranteed performance and is broadly applicable to other safety critical decision-making scenarios.
Oct 27, 2026

Conservative Policy Iteration For Zero-Sum POSGs

Occupancy-state reductions make dynamic programming possible for zero-sum partially observable stochastic games (zs-POSGs). Existing point-based solvers usually optimise a central value function. This restores Bellman recursion, but it requires point-based updates over high-dimensional occupancy spaces. We introduce an approximate policy-iteration algorithm for finite-horizon zs-POSGs. The method maintains one policy per player and updates both jointly along occupancy trajectories induced by conservative policy updates. Instead of approximating a full central value function, each update uses a single supporting hyperplane induced by the current joint policy. The algorithm alternates local policy improvements, conservative updates, and persistent exploration. Computation is therefore concentrated on occupancy regions where the current joint policy is exploitable. On finite-horizon zero-sum POSG benchmarks, the method is competitive at small horizons and often improves the runtime-exploitability trade-off at larger horizons compared with PBVI, HSVI, and CFR+.
Oct 26, 2026

PolicyMediator: Multi-Regional Policy Learning through Mediation Agents with Evolving Skills in OPC Ecosystems

One-person companies (OPCs) are emerging enterprises operated by individual entrepreneurs through digital platforms and LLM-based AI agents. These highly skilled entrepreneurs are characterized by being digital nomads, with limited reliance on local ecosystems. Governments in China use distinct regional support policies to attract mobile OPCs, turning local support into a cross-regional coordination problem. OPCs compare regions to maximize their own utility, but imperfect policy perceptions may cause them to reject otherwise suitable locations. Existing policy-learning methods optimize resources under fixed perceptions, while consensus-formation methods aggregate views without feeding cognitive changes back into policy learning. This disconnect produces mismatched location choices and inefficient regional policy adjustment. We introduce PolicyMediator, a cognition-constrained leader-multi-follower Stackelberg framework with bilateral mediation. The mediator explains verified policies to OPCs to build awareness and trust, and reports their needs and unresolved concerns to governments for policy learning. Its self-evolving skills retain, revise, or prune strategies using verification, allocation, capacity, and objective trajectories rather than mediation success alone. Across six-region simulations in China with five OPC types, five scenarios, and three OPC-simulator backbones, PolicyMediator ranks first in all 15 OPC-backbone--scenario combinations. Over the eight-baseline mean, it improves the aggregate government objective, combining economic performance, employment benefit, and policy-access equality, by 42.4% and average OPC utility by 40.9%.
Oct 26, 2026

Multiscale Policy Optimization for Long-Horizon LLM Agents

Long-horizon LLM agents interact with their environments through complete assistant turns while generating each action token by token. We observe that critic values vary primarily across turns, motivating optimization anchored at turns with finer credit allocation among tokens. We propose Multiscale Policy Optimization (MPO), combining Multiscale Advantage Estimation (MAE) and turn-level importance sampling to align credit assignment and policy updates at the turn scale. Extensive experiments across search, tool-use, and automation tasks with 4B and 30B-A3B models demonstrate consistent gains over strong RL baselines.
Oct 23, 2026

When to Stop Penalizing Actions? Martingale GRPO for VLA Reinforcement Learning

A failed vision-language-action (VLA) rollout does not imply that every action contributed to failure. Yet in group-relative policy optimization (GRPO), a single trajectory-level advantage is assigned to all actions in a rollout. Potentially useful recovery attempts can thus be penalized alongside the errors they aim to correct. Ideal stopping conditions for failure feedback are characterized using success-probability and policy-score martingales. At these stops, the removed gradient contribution has zero conditional expectation given the retained information. Truncation therefore preserves the expectation of a reference GRPO group-gradient estimator and can reduce its covariance. One such stop occurs when the conditional probability of success is zero, motivating error and recovery cues for selecting candidate feedback boundaries. Martingale GRPO (MART-GRPO) is introduced with a vision-language model (VLM) for event-guided truncation. The VLM is fine-tuned on our annotated training set using low-rank adaptation (LoRA). The last visible error without observed recovery is localized by the frozen VLM in complete rollout videos. Failure feedback is retained through the selected error interval and removed afterward, while successful-rollout loss terms and original loss normalization are preserved. For retrospective boundaries, changes in expectation and variance are analyzed to assess departures from ideal stopping. Experiments with OpenVLA-OFT on RoboTwin 2.0 show that MART-GRPO improves success rates over full-trajectory GRPO across multiple tasks.
Oct 23, 2026

Dual-Adaptive Z-Score PPO: Dynamic Clipping Using Action Confidence and Normalized Advantage Magnitude

Proximal Policy Optimization (PPO) remains a popular on-policy reinforcement learning algorithm owing to its stability and ease of implementation. Its clipping threshold, however, is fixed, so every training sample receives the same trust-region width regardless of how reliable the corresponding policy update is. This work examines adaptive-clipping variants of PPO in which the constant clipping parameter is replaced by sample-dependent thresholds computed from policy and return statistics. The central method, Dual-Adaptive Z-Score PPO (PPO-DAZ3), sets the clipping interval for each sample from two signals: action confidence, given by the probability of the selected action under the current policy, and update confidence, given by the normalized advantage magnitude. The rest of the standard PPO training pipeline is left unchanged. Evaluation on three classic control benchmarks (LunarLander-v3, CartPole-v1, and Acrobot-v1) indicates that adaptive clipping improves learning in some settings but not all. PPO-DAZ3 performs best on LunarLander-v3, the smooth advantage-based variant PPO-AZ3S leads on Acrobot-v1, and standard PPO stays competitive on CartPole-v1, suggesting that adaptive clipping is most beneficial on harder control tasks.
Oct 22, 2026

MOER: Mix-Policy Optimization under Expert Prefix Guidance with Selective Rollback

Mix-policy optimization incorporates off-policy expert data into the reinforcement learning (RL) loop to transcend the capability boundary of on-policy RL. However, existing methods either provide static full-trace guidance that suppresses self-exploration, or adopt expert prefix guidance without correctness guarantees, leading to suboptimal performance or even training collapse. In this paper, we propose MOER, a mix-policy optimization framework that harmonizes expert guidance with self-exploration. MOER injects a decaying expert chain-of-thought prefix into the prompt and lets the model complete the remaining reasoning, with the prefix length progressively annealed to encourage self-exploration. A selective rollback mechanism replaces failed completions with the original full expert trace, guaranteeing the correctness of every off-policy trajectory. Furthermore, a harmonized mix-policy objective optimizes on-policy tokens with a GRPO loss while supervising both expert prefix and completion tokens with a trust-region SFT loss, substantially stabilizing training. Extensive experiments show that MOER achieves the highest average scores in mathematical and general domain reasoning tasks, outperforming all basic and mix-policy baselines, and generalizes consistently across diverse tasks, model scales, and base model categories. Codes are available at https://github.com/CityU-MLO/MOER.
Oct 21, 2026

One Act, Many Masks: Effect-Marginal Policy Optimization for Long-Horizon Agents

Recent advances in multi-turn reinforcement learning (RL) have substantially improved the performance of LLM agents on complex, long-horizon interactive tasks. However, despite progress in fine-grained credit assignment, policy optimization still applies these learning signals through response-specific probability ratios. Different textual responses can induce the same executable effect, yet credit differences arising from subsequent outcomes remain tied to their individual textual realizations. We identify this mismatch as credit-update misalignment. Our analysis reveals that credit-text pairing directly shapes policy-update directions, even when the policy context and executable effect are held fixed. Motivated by this finding, we propose Effect-Marginal Policy Optimization (EMPO), which aligns policy optimization with execution-level equivalence through an effect-level surrogate that couples updates across alternative textual realizations. EMPO preserves the original credit signals, integrates with both trajectory-level and step-level credit estimators, and requires no additional critics or rollouts. Experiments on ALFWorld, WebShop, and Sokoban show that EMPO outperforms strong baselines in both task performance and learning efficiency.
Oct 20, 2026

Doob-Based Generative Policy Optimization

Diffusion and flow models provide expressive policy classes for robotic and sequential decision-making systems. However, their action likelihoods are generally intractable, limiting the direct use of widely adopted policy optimization methods such as PPO and GRPO. We propose Doob Policy Optimization (DoobPO), a framework that leverages Doob's hh-transform for likelihood-free reinforcement learning fine-tuning of generative policies. DoobPO separates policy improvement from generative policy learning. It first learns a lightweight ratio model that represents the desired policy update, and then distills the resulting ratio-tilted target policy into the generative policy via conditional drift matching. This formulation applies uniformly to both diffusion and flow policies and supports a broad family of policy improvement rules, including PPO and GRPO. We prove that the practical conditional drift matching objective is gradient-equivalent to the ideal Doob drift matching objective. We further show that the induced diffusion process solves an endpoint-constrained stochastic optimal control problem. This establishes that it achieves the minimum-energy modification required to realize the desired policy tilt. Across MuJoCo continuous-control benchmarks, DoobPO achieves strong performance against competitive diffusion- and flow-policy baselines, showing a principled and practical route to RL fine-tuning for generative policies.
Oct 17, 2026

Bellman Language Models: Solving the Recurring Decisions of LLM Agents

A general observation in agentic workflows, especially bounded ones, is that the decision space of the LLM, when we separate a decision from generation, starts approaching saturation: the agent meets the same states again and again, and after enough conversations it stops meeting new ones. LLMs generally don't treat these decisions as an optimization problem and instead re-derive them from their context/prompt on every turn. We show that treating decision making within reasonable bounds as an optimization problem, solving it once, and reusing the solved decisions can often make an agent faster, cheaper and more reliable. Our approach, which we call Discretize, Then Decide, starts from data such as the workflow's own logs or, when none exists, runs of the LLM agent itself. It reads each conversation into discrete states and estimates the probability of each outcome and the cost of each move, until no new states appear and the probabilities stop changing. It then solves the resulting Markov decision process. The solution runs as a controller around the model, sending moves whose content is fixed in the state, such as a tool call or a templated reply, and any state it hasn't seen is sent back to the LLM. We call an LLM run this way a Bellman language model (BLM), after Richard Bellman's equation for the best move in each state. On τ2τ^2-bench telecom, the controller with no model solves 79% of tasks. As a BLM, Qwen3.8-27B, an open model, goes from 29% to 96% with 93% fewer model calls. For Claude Opus 5.5, which solves 96% of tasks on its own, running it as a BLM cuts model calls and cost per task by about 98% and lets it solve 2 more tasks. On InventoryBench, the solved policy with no model beats every published method. Where the hard part is writing content rather than choosing moves, such as resolving SymPy issues, few decisions can be solved in advance, and a BLM helps little.
Oct 14, 2026

Lifted State Policy Optimization: Internalizing Self-Correction through Transition Geometry

Preference optimization ranks final answers and trajectory RL maximizes aggregate returns, yet neither credits the local change that turns a flawed reasoning state into a better one. We introduce Lifted State Policy Optimization (LSPO), which establishes this finite, cost-adjusted transition as the primary unit of learning. Each textual answer is augmented with a learned continuous coordinate, forming a lifted state that an energy function evaluates and orders. A shaped reward drives energy descent net of edit and step costs, and an internalization loss absorbs selected low-energy answers into the direct generator. On Qwen3.5-27B, the resulting single-pass policy reaches 71.8% aggregate accuracy across mathematics, science, code, and logic benchmarks, matching test-time reflection at 71.9% with 28% lower latency and 25.8% fewer tokens. The pattern holds across dense and mixture-of-experts backbones, and matched controls confirm that transition credit and the lifted coordinate each contribute, as does answer internalization. Our code is available at https://github.com/Venomrko/lspo.
Oct 12, 2026

DR-APO: Decoupled-Representation Advantage-Guided Policy Optimization from Continuous Control to Embodied Navigation

Policy representations in reinforcement learning have expanded beyond Gaussian policies to diffusion and flow-matching based generative policies. However, existing RL methods remain representation-specific, lacking a unified optimization interface for emerging generative policies. In particular, extending likelihood-based objectives to multi-step generative policies couples optimization to their internal generation process, where accumulated likelihood ratios can weaken the update signal. Meanwhile, generative-policy-specific methods often rely on multi-candidate generation and critic-based selection, improving overall return while providing limited improvement to the policy itself. We introduce DR-APO (Decoupled-Representation Advantage-Guided Policy Optimization), which decouples policy improvement from policy representation through advantage-guided target improvement and representation-compatible direct action fitting, providing a shared optimization principle for Gaussian, diffusion, and flow policies. Across five MuJoCo-v5 tasks, DR-APO remains competitive with SAC for Gaussian policies and achieves the best aggregate policy-only performance for both diffusion and flow-matching policies. NavDP post-training improves the Success Rate (SR) from 70.33% to 89.55% in simulation and from 25% to 65% in real-world navigation. These results demonstrate a unified policy-improvement principle across Gaussian, diffusion, and flow policies, from continuous-control learning to embodied-policy post-training.
Oct 9, 2026

Generalized Proximal Policy Optimization: Unifying LLM RL Algorithms from the Perspective of Divergence

Reinforcement learning has become the foundational paradigm for enhancing the reasoning capabilities of LLMs. Yet the dominant approach, Proximal Policy Optimization (PPO), relies on a geometrically degenerate clipping proxy that cannot adapt its resistance to policy shifts, leading to premature clipping and degraded exploration. Subsequent variants alleviate these issues through fragmented and heuristic modifications. We propose Generalized Proximal Policy Optimization (GPPO), a principled and unified framework for understanding and designing LLM RL algorithms through the lens of divergence. We establish a generalized ff-divergence trust-region theory with a monotonic performance improvement guarantee, and show that different ff-divergences induce distinct geometries and policy update dynamics. We further introduce novel estimator topologies that provide more faithful divergence estimates than the conventional single-sample Monte Carlo estimation. Under this framework, PPO and recent variants such as DAPO, DPPO, and RIPO emerge as special cases, establishing a principled two-dimensional design space of {Divergence ×\times Estimator}. Extensive experiments on eight competition-level math and coding benchmarks show that both dimensions are critical. The best-performing GPPO instantiation surpasses standard GRPO by up to 50% and 14% on math and coding, delivering SOTA performance.
Oct 9, 2026

Cross-Behavior Policy Optimization: Unified Learning from Offline Demonstrations and On- and Off-Policy Rollouts via a Shared Regression Target

Reasoning post-training draws on offline demonstrations and policy-generated rollouts, yet conventional supervised fine-tuning (SFT) followed by reinforcement learning (RL) separates their objectives and limits continued reuse of collected responses. We propose Cross-Behavior Policy Optimization (CBPO), which makes these sources compatible with one regression loss through a fixed-reference target. Because the generating policy does not enter this target, positive and negative offline completions, reference responses, and current and historical rollouts can share a common reward-labeled record format and replay buffer. As new responses enter the buffer, optimization retains the same regression objective without coordinating separate SFT and RL objectives or requiring stored behavior-policy probabilities, importance weighting, or a learned critic. Restricting the available sources gives offline, on-policy, and off-policy variants, while CBPO-Unified combines them. Across two model scales and four mathematical reasoning benchmarks, CBPO delivers leading avg@32 and pass@32 performance across benchmarks and data regimes. On Qwen3-1.7B, CBPO-Unified improves benchmark-averaged avg@32 and pass@32 over SFT DAPO by 2.33 and 4.44 percentage points, respectively. In the replay ablation, retaining historical rollouts improves these metrics by 1.87 and 3.72 percentage points over the no-history configuration.
Oct 8, 2026

Closing the Horizon Gap in Policy Optimization for Adversarial MDPs

We consider policy optimization for online episodic tabular Markov decision processes (MDPs) with adversarial losses and bandit feedback. Policy optimization updates the policy locally at each state and avoids optimization over the occupancy-measure polytope, but its existing regret bounds are larger by a factor of the horizon HH than those of occupancy-measure-based algorithms. We close this gap by using regularized QQ-functions, which allow us to control the stability of the local updates jointly over all state-action pairs rather than separately at each state. The resulting algorithm attains high-probability regret bounds of O~(HS(H+A)T)\widetilde O(\sqrt{HS(H+A)T}) for known transitions and O~(HSAT)\widetilde O(HS\sqrt{AT}) for unknown transitions, where SS is the number of states, AA the number of actions, and TT the number of episodes. Both bounds improve the horizon dependence of existing policy optimization bounds, and the latter matches the best-known bound. We further extend the algorithm to adversarial linear-mixture MDPs and obtain the same improvement in the horizon dependence.
Oct 4, 2026

Pessimistic Minimax Learning for Public-Private Information Games under Unilateral Coverage

We study offline learning in two-player zero-sum contextual games with public and private information, motivated by strategic settings such as auctions and negotiations with private valuations. We introduce unilateral prescriptive concentrability and show that asymmetric information can change offline coverage through its effect on equilibrium behavior. For finite state-action spaces, we develop a pessimistic algorithm with an O~(1/n)\tilde{O}(1/\sqrt{n}) exploitability rate, matching the standard sample-size dependence for fully observed minimax games. We further develop a pessimistic policy mirror descent framework, PPA-PMD, for general function approximation and obtain a unified O~(1/n+1/T)\tilde{O}(1/\sqrt{n} + 1/\sqrt{T}) exploitability rate with no-regret actor updates. Together, these results provide the first theoretical framework for offline equilibrium learning under public-private information constraints.
Oct 2, 2026

Sharp Convergence and Sample Complexity of Policy Mirror Descent for Average-Reward MDPs

Policy mirror descent (PMD) has a mature finite-time theory in discounted Markov decision processes (MDPs), but less is known in the average-reward setting, a more natural objective for many control applications. We give a finite-time, finite-sample analysis of PMD in ergodic average-reward MDPs built around a single master recursion that governs convergence for any critic, without external regularization. Its specializations yield linear rates for exact, inexact-tabular, and linear function approximation (LFA) updates. For exact PMD, the rate is tight in the worst case, and convergence can become superlinear. We complement these convergence results with critics whose end-to-end sample complexities improve the prior best policy-gradient dependence from ∣S∣∣A∣ tmix3|S||A|\,t_{\mathrm{mix}}^3 to ∣S∣ tmix3+∣S∣∣A∣ tmix2|S|\,t_{\mathrm{mix}}^3+|S||A|\,t_{\mathrm{mix}}^2 in the tabular setting and from ∣S∣∣A∣ tmix5|S||A|\,t_{\mathrm{mix}}^5 to d tmix3d\,t_{\mathrm{mix}}^3 under LFA. Matching information-theoretic lower bounds establish that both critics are optimal up to logarithmic factors.
Oct 2, 2026

On the Convergence of Success Conditioning for Policy Optimization

Success conditioning is a strategy for improving decision-making policies in stochastic environments; it updates a policy by increasing the probability of taking actions that yield successful outcomes. Success conditioning is common to many reinforcement learning applications, yet its limiting behavior and convergence rates are not well understood. In this work, we demonstrate that success conditioning converges to an optimal policy on a broad class of Markov decision processes (MDPs). We also derive convergence rates in some common settings. For discounted MDPs, we prove convergence within O(1/εp)\mathcal{O}(1/\varepsilon^p) iterations to an ε\varepsilon-optimal policy, where the exponent pp depends on problem data. For single-period MDPs, such a policy is obtained within O(log⁡(1/ε))\mathcal{O}(\log(1/\varepsilon)) iterations.
Oct 2, 2026

Sample complexity of variance-reduced policy gradient: weaker assumptions and lower bounds

Several variance-reduced versions of REINFORCE based on importance sampling achieve an improved O(ε−3)O(ε^{-3}) sample complexity to find an εε-stationary point, under an unrealistic assumption on the variance of the importance weights. In this paper, we propose the \algo (Defensive Policy Gradient) algorithm, based on defensive importance sampling, which achieves the same rate without any assumption on the variance of ordinary importance weights. We also establish lower bounds in a generalized black-box policy-optimization model that hides states and actions and permits parameter-dependent rewards. In this model, the optimal rates are Θ(ε−4)Θ(ε^{-4}) with bounded-variance one-policy feedback and Θ(ε−3)Θ(ε^{-3}) with mean-square-smooth coupled two-policy feedback. Under standard policy-regularity conditions, REINFORCE and \algo realize the corresponding oracle conditions and attain the O(ε−4)O(ε^{-4}) and O(ε−3)O(ε^{-3}) upper bounds, respectively. Although the lower bounds do not apply directly to the classical MDP interaction model in which these algorithms operate, this correspondence provides oracle-level evidence that the faster rate of \algo is optimal and genuinely separated from that of vanilla policy gradient.
Oct 1, 2026

FERPO: Forward Entropy-Regularized Policy Optimization

Several state-of-the-art methods for online reinforcement learning in continuous control improve policies using action gradients of a learned critic. However, critics are typically trained to predict returns, and accurate value predictions do not necessarily yield accurate action derivatives, potentially leading to unreliable policy updates. We propose Forward Entropy-Regularized Policy Optimization (FERPO), an on-policy maximum entropy reinforcement learning algorithm that performs policy improvement using critic values without differentiating the critic with respect to actions. FERPO derives an optimal target action distribution from a policy-improvement objective regularized by entropy and Kullback-Leibler (KL) divergence. We then fit the actor to this target by minimizing a forward-KL objective, estimated using self-normalized importance sampling (SNIS) with actions drawn from the rollout policy. By limiting the target distribution's deviation from the rollout policy, the KL regularization helps keep these importance weights well behaved. In contrast to reverse-KL objectives, which can favor a subset of the target distribution's modes, the forward-KL objective encourages coverage of multiple high-value modes and thereby promotes exploration. Experiments and ablations on MuJoCo Playground and ManiSkill show competitive performance and sample-efficiency gains. Computational benchmarks also demonstrate faster actor updates than Relative Entropy Pathwise Policy Optimization (REPPO).
Oct 1, 2026

Iterative Policy Refinement through Semantic Rollout Analysis

Structured policies improve efficiency, robustness, and interpretability in imitation learning by introducing task-specific inductive bias, but existing structure generation methods rely either on extensive human input or on static domain knowledge encoded in LLMs, which may be inconsistent with the expert demonstrations. We propose a closed-loop framework that iteratively refines structured policies using LLM-guided analysis of policy rollouts. By logging rollouts as semantically meaningful tabular data and prompting the LLM to generate diagnostic analysis code, our method identifies suboptimalities in the policy structure and iteratively corrects them without requiring human instruction. Experiments on car racing and door opening tasks show that our approach improves imitation learning performance by up to 15% over zero-shot LLM-generated structures and requires 75% less compute to achieve the same reinforcement learning performance. These results demonstrate that tabular rollout analysis provides an effective feedback signal to align LLM-generated policy structures with expert demonstrations, and we can utilize it to generate good policy structures automatically.