Policy Learning

Reinforcement learning with verifiable rewards (RLVR) often adopts GRPO-style group-relative updates, sampling multiple rollouts per prompt to construct normalized learning signals. However, merely increasing the number of rollouts does not reliably strengthen learning: under GRPO-style group normalization, per-rollout policy-gradient features can concentrate into a low-rank, signed geometry, causing substantial cancellation during aggregation and weakening the effective update. We address this failure mode with SALT, a Subspace-Adaptive geometry pLug-in componenT that uses sample-wise gradient geometry to reweight the coefficients of group-relative updates. SALT estimates a dominant shared subspace from the mini-batch Gram geometry, decomposes group-relative coefficients into shared and residual channels, and adaptively amplifies the residual channel when signed cancellation is severe. Across diverse reasoning-oriented RLVR benchmarks and model scales, SALT improves effective update geometry and performance without modifying the reward model or the rollout sampling procedure

LLM agents operate in two distinct regimes: open-weight agents amenable to reinforcement learning (RL) and black-box agents whose behaviour must be controlled purely at test time. Although black-box agents are often backed by state-of-the-art proprietary LLMs, API-only access precludes parameter-level optimization, rendering most RL methods inapplicable. To address this limitation, we turn to a known equivalence between RL and Bayesian inference. We propose Agentic Monte Carlo (AMC) to directly sample from the optimal policy of a black-box agent rather than training it through RL. The optimal policy is a posterior over trajectories whose prior we define as the fixed black-box LLM agent. We employ Sequential Monte Carlo to sample from this posterior by learning a value function to steer the agent while leaving the underlying black-box model unchanged. We validate AMC on three diverse environments from the AgentGym benchmark, demonstrating significant improvements over prompting baselines and even outperforming Group Relative Policy Optimization (GRPO) as we scale the test-time compute of our method. AMC demonstrates the feasibility of performing principled RL-style optimization of black-box LLM agents. Code is available at https://github.com/layer6ai-labs/Agentic-Monte-Carlo

Reasoning models have advanced rapidly, but the dominant reinforcement learning from verifiable rewards (RLVR) recipe remains surprisingly narrow: sample many responses and reward each with a single bit indicating whether the final answer is correct. Yet many settings provide rich feedback, including execution traces, tool outputs, expert corrections, and model self-evaluations. We study how to use such feedback through a distributional variant of the classic imitation learning algorithm DAgger, where the learner has local access to an expert distribution on states visited by the current policy. This yields a simple forward cross-entropy objective that admits a blackbox expert and whose sequence-level gradient {conduct rich credit assignment by propagating} future expert-student disagreement back to earlier decisions. We show that prior RL with self-distillation objectives based on reverse KL or Jensen-Shannon fail to guarantee monotonic policy improvement: even when the expert has higher reward, their updates may increase probability on worse actions. In contrast, we show that forward cross-entropy admits monotonic policy improvement and enjoys guarantees on regret. We further show that our objective optimizes a lower bound on teacher-weighted likelihood of success, leading to improved Pass@N. Empirically, our approach, DistIL, improves over RLVR and RL with self-distillation baselines across a variety of domains: scientific reasoning, coding, and solving hard mathematical problems.

Reinforcement learning (RL) for large language models usually supervises reasoning with scalar outcome rewards, such as binary correctness. Such rewards provide an optimization direction but rarely explain how a model should revise its mistaken reasoning, which can encourage shortcut learning and brittle policies. We propose \textbf{SocraticPO} (Socratic Policy Optimization), a policy-optimization framework that augments RL rollouts with Socratic-style natural-language guidance. During rollout, the student first answers independently; if the answer is incorrect, a teacher diagnoses the attempt and provides concise corrective guidance, after which the student continues under the expanded context. Crucially, this guidance is paired with reward decay: correct answers obtained after teacher intervention only receive decayed rewards, preventing the policy from treating teacher help as a free path to reward. Since SocraticPO only modifies the rollout process while leaving the standard expected-reward objective intact, it can be plugged into existing policy-gradient backends such as Reinforce++. Moreover, because the teacher provides only text-level guidance, SocraticPO can leverage stronger black-box teacher models without requiring access to logits or distribution matching. On undergraduate-level scientific reasoning benchmarks from SciKnowEval, SocraticPO improves over strong RL and self-distillation baselines. Ablations show that both targeted guidance and reward decay are necessary, with reward decay mitigating reliance on assisted correction.

Reinforcement learning from verifiable rewards with GRPO is a standard approach for post-training reasoning LLMs. It remains sample inefficient. Each rollout is used for a single gradient update and then discarded. Naive replay is not well suited in this setting because LLM policies drift quickly per gradient step. Stored rollouts therefore become stale and can destabilize training. We propose a rollout-level replay buffer for GRPO that stores and samples individual rollouts rather than whole groups. The buffer bounds staleness through age eviction. Any rollout older than tau_max training steps is removed. The buffer also preserves on-policy data via fresh-anchored composition. Each batch keeps its fresh on-policy rollouts and then concatenates replay rollouts drawn separately from the buffer. We prioritize replay by per-rollout advantage magnitude and recycle individual rollouts whose advantages are large. Across three Qwen3-Base scales on five math benchmarks, our method outperforms GRPO and naive replay baselines. Gains are positive at every scale and reach +1.66 pp on the five-benchmark average at 4B. Under an AES metric that jointly measures accuracy and token efficiency, our method is the only condition with a positive margin over GRPO at every scale.

The framework of robust Markov decision processes (RMDPs) allows the design of reinforcement learning agents that satisfy performance guarantees under worst-case transition dynamics. Traditional RMDPs consider discrete-time dynamics and recently, sample-efficient policy gradient algorithms have been considered in this context. This paper investigates policy gradient algorithms within a continuous-time RMDP framework. Policy gradients and adversarial gradients are derived using pathwise and adjoint-based formulas for stochastic and ordinary differential equations. We propose double-loop optimisers to obtain linear convergence in the oracle-based setting and an $\tilde{\mathcal{O}}(\frac{1}{ε^2})$ sample complexity in the sample-based setting in an analysis which also derives novel tools for the framework of undiscounted total cost MDPs. Additionally, we propose mean-field optimisers as distributional optimisers with an $\tilde{\mathcal{O}}(\frac{1}{K})$ oracle-based convergence rate and an $\tilde{\mathcal{O}}(\frac{N^2}ε)$ sample complexity under $N$-particle approximation. The effectiveness of continuous-time policy gradient algorithms is confirmed for both optimisers on continuous-time RMDPs with neural ordinary differential equation dynamics.

Ultra-high-resolution image sensors offer the potential to capture fine spatial details critical for many visual perception tasks, but acquiring and processing all pixels at full resolution is often infeasible under realistic bandwidth, latency, and power constraints. Existing approaches address this challenge through acquisition strategies such as spatial or temporal downsampling, which irrevocably discard information before task relevance can be assessed. In this work, we introduce a real-time, predictive, and task-aware foveated imaging system that operates directly at image acquisition time. Leveraging emerging dual-stream sensor architectures, our method dynamically allocates limited pixel bandwidth to task-relevant regions of interest while maintaining a low-resolution global context. We formulate foveated acquisition as a sensor attention policy-learning problem, in which past observations guide actions that determine future measurements, closing the perception-acquisition loop. Through extensive simulation across multiple perception tasks, we demonstrate that our approach achieves high task performance under strict pixel budgets and significantly outperforms relevant baselines operating at the same bandwidth. We further validate our system on a 200-megapixel dual-stream sensor, capturing real-world videos under realistic bandwidth and latency constraints, demonstrating the practical feasibility of task-driven, acquisition-time foveated imaging.

Distilling expert demonstration data into large generative models using behavioral cloning is a scalable approach to learning capable policies for robotic control, particularly for dexterous manipulation. Reinforcement learning (RL) can be used as a means to finetune these policies further using additional experience. An open question is whether RL is more sample-efficient than collecting more human demonstrations. Prior work has finetuned large pretrained policies in a scalable fashion by applying RL to a smaller residual policy that corrects the pretrained model. However, for the typical sparse reward tasks, RL algorithms can struggle to optimize the behavior in a sample-efficient manner. We explore inverse reinforcement learning, where a dense reward function is learned from expert demonstrations, potentially reducing the challenge of RL finetuning. We specifically consider coherent imitation learning, an IRL method that facilitates improvement of the BC policy through using a specific reward formulation with theoretical guarantees. We show that our IRL method maintains or improves the performance of pi-0.5 on all six sparse manipulation tasks and achieves a $\geq 90\%$ success rate on five out of six complex manipulation tasks, outperforming RL-based baselines using sparse rewards. By ensuring our initial pretrained finetuning policy is optimal for our initial reward and critic, our method circumvents the initial drop commonly seen in RL finetuning and enables faster improvement.

Robot learning must produce policies that generalize to new combinations of constraints, teammates, and environments. To achieve this, we must structurally factor the policy, which is a choice that dictates what generalizes, what requires retraining, and what remains entangled. Existing methods span a wide spectrum, from expecting structure to emerge from data scaling, to hand-designing it via hierarchies, skill libraries or learned specializations. In this paper, we study what we argue is the most fundamental factorization in robotics: separating the world from the task. We investigate the conditions under which this factorization is principled. World factors are properties of the embodied system and the environment; they exist independently of intent. Task factors are defined by the task's logic over what the world admits. We formalize this asymmetry through Bayesian model evidence: it aligns with the data-generating process, maintains high likelihood through an analytical world model, and reduces the Occam razor's penalty on task parameters. We instantiate this factorization by pairing AICON, a differentiable graph of recursive estimators and interconnections that is compositional, operates without task-specific data, and propagates cost gradients to actuators, with a compact, learned policy that modulates gradient paths. Gradients serve as the interface between the two factors: they carry world structure through the graph and task structure through costs, enabling low-dimensional learning while preserving structural generalization. We test the world/task factorization across three problems that encompass heterogeneous robots, environments, task logic and sensorimotor modalities. Our framework outperforms end-to-end baselines and analytical heuristics in all settings, generalizes zero-shot to out-of-distribution configurations, and transfers to real hardware without retraining.

While diffusion-based policies have impressive performance and expressivity, their long offline training slows down the data collection and policy deployment loop. We introduce Closed-Form Diffusion Policies, a class of training-free diffusion-based policies for imitation learning using the closed-form score derived from the demonstration dataset. We deploy CFDP with real-time inference with a mobile CPU in hardware experiments, showing it can successfully perform imitation directly from the dataset in milliseconds and with faster inference than neural diffusion policies. In experiments on imitation learning benchmarks, we show that CFDP is competitive against neural baselines that require hours of training, providing a favorable tradeoff between training time and performance. Finally, we show how closed-form diffusion policies act as a composable primitive that enables data-driven inference-time editing of pre-trained neural diffusion policies, including policy guidance and novel demonstration augmentation.

Behavior cloning with high-capacity generative policies achieves strong imitation performance, but is often limited by demonstration coverage and distribution shift. Direct reinforcement learning fine-tuning can improve performance, but updating large action decoders is frequently unstable and sample inefficient. We propose Lagrangian Perturbation Diffusion Steering (LP-DS), a lightweight adaptation method that improves a frozen generative policy by learning a compact noise-space perturbation before decoding. LP-DS optimizes this perturbation with a Lagrangian trust-region objective, improving downstream value while constraining deviation from the latent prior. Across RoboMimic manipulation, OpenAI Gym locomotion, and Adroit dexterous manipulation benchmarks, LP-DS improves sample efficiency, success, and return while maintaining higher action-space entropy than unconstrained noise-space steering, with return improvements of up to 25% over prior baselines. Additional evaluations with flow-matching backbones, a large vision-language-action model, and physical Franka deployment show that LP-DS is not limited to compact diffusion policies or simulated benchmarks. Project page: https://sites.google.com/view/lp-ds/home.

Decision-focused learning (DFL) trains predictive models by optimizing downstream decision quality rather than standalone prediction accuracy. For contextual linear optimization, most existing DFL methods assume offline data and full observations of the objective cost vector. We develop an on-policy learning method for sequential contextual linear optimization under partial feedback, generalizing the standard bandit feedback setting. Our method learns a stochastic predict-then-optimize policy that samples a cost-vector prediction from a conditional distribution and solves the resulting downstream linear optimization problem. To update this distributional model, we introduce a two-component hybrid gradient estimator. The first component is a score function estimator, which provides an unbiased but potentially high-variance policy gradient estimate. The second is a decision-focused plug-in component that uses an auxiliary nuisance estimate of the latent cost vector to exploit the downstream optimization structure, becoming more informative as the estimate improves. We prove an O(T^-1/2) bound on the average squared policy-gradient norm, matching the standard non-convex SGD rate. Experiments on top-k selection, shortest path, combinatorial pricing, and a real-data energy-scheduling benchmark show that, under bandit feedback, the hybrid gradient approach achieves lower mean cumulative regret than the contextual-bandit baselines on all four benchmarks and than linear Thompson sampling on three, and that it also works with richer conditional generative cost models. Code is available at https://github.com/Joeyetinghan/on-policy-bandit-dfl.

Inductive generalization is a framework for reinforcement learning (RL) generalization in which inductively related task instances admit inductively related policies. Prior work captures this structure via a higher-order policy-evolution function learned directly with RL, but suffers from poor training scalability: as training tasks grow, aggregated reward feedback becomes noisy and conflicting, destabilizing training and weakening generalization. We propose DIBS, a decoupled behavioral cloning approach that separates learning task-specific policies from learning the evolution function. We first learn individual teacher policies per task via standard RL, then fit the evolution function via behavioral cloning on teacher-labeled state-action pairs. This replaces noisy reward aggregation with dense, stable supervision. DIBS achieves significant improvements in both training stability and zero-shot generalization against existing RL and meta-RL algorithms.

Offline reinforcement learning requires improving a policy from fixed data while avoiding out-of-distribution actions with unreliable value estimates. Diffusion and flow policies handle this trade-off by modeling the behavior distribution to regularize the RL objective, but they require iterative denoising, solver integrations, and in more efficient variants, distillation or other approximations at inference. We propose DriftQL, which combines a drift-based behavioral regularizer with critic-driven policy improvement. The value signal biases the policy toward high-value regions of the data support, while attraction and repulsion together keep generated actions near the data and prevent collapse onto a single mode. DriftQL is implemented as a single network with a unified training objective and generates actions in a single forward pass. On D4RL and OGBench, DriftQL consistently outperforms diffusion and flow methods, advancing the state of the art. Under degraded data quality, where the baselines visibly struggle, DriftQL remains close to its clean-data performance, positioning it as a promising alternative to diffusion and flow-based methods while maintaining the simplicity and efficiency of deterministic approaches. Project page: https://driftql.github.io/

We propose Parameterized Diffusion Policy (PDP), a framework for learning diffusion policies conditioned on low-dimensional, continuous parameters embedded in a learned behavior manifold. By constructing this manifold so that distances between latent representations reflect the semantic similarity between physical trajectories, we transform diffusion from a mechanism for stochastic diversity into a precise and optimizable tool for behavior steering. Our approach enables smooth interpolation between known strategies and efficient adaptation to novel constraints without updating policy weights. We demonstrate that PDP significantly improves adaptation performance on complex multimodal benchmarks in both simulated and real-robot experiments compared to standard diffusion policies, particularly in scenarios requiring the synthesis of novel behaviors.

Large language model agents trained with reinforcement learning (RL) often learn brittle, task-specific shortcuts. We hypothesize that agents generalize better when their successful trajectories are structurally compressible, decomposed into a small set of reusable abstract patterns. To formalize this, we introduce ReuseRL, which grounds agentic RL in the Minimum Description Length (MDL) principle. ReuseRL extracts a shared skill dictionary from successful trajectories and augments the RL objective with a segmentation cost, explicitly penalizing idiosyncratic behaviors that encode poorly. We prove a PAC-Bayes bound guaranteeing that a dictionary extracted from successful trajectories has bounded expected description length on future successful behavior. Across ALFWorld, TextWorld-Cooking, and Countdown-Stepwise, ReuseRL improves in- and out-of-distribution success over vanilla GRPO and strong round-length baselines.

Tree search is a central abstraction behind many language-agent reasoning and decision-making tasks: agents must explore actions, remember failures, and backtrack toward promising alternatives. Yet, we lack a theoretical understanding of how transformer-based policies acquire such search capabilities from the training dynamics of reinforcement learning (RL). We study this question in a stochastic $k$-ary tree environment, where an agentic transformer observes only its trajectory history through interaction and receives a terminal reward for reaching a hidden leaf goal node. We first construct a two-head transformer that implements randomized depth-first search (DFS): one head tracks previous actions, while the other detects failure outcomes and triggers backtracking. We then analyze the training dynamics of policy gradient under a depth-wise curriculum, showing that this same DFS mechanism emerges in stages from sparse reinforcement feedback without expert demonstrations. The resulting policy exhibits depth generalization: after training only on depth-$1$ and depth-$2$ trees, it succeeds on deeper full trees. We further show that, under imbalanced goal distributions, discounting the return leads to a ranked DFS policy that prioritizes higher-probability branches. Overall, our results identify a mechanistic normal form for transformer-based search, in which attention heads specialize and cooperate to extract decision-relevant traces from context and convert them into agentic action selection via RL training.

On-policy distillation (OPD) trains a student on prefixes sampled from its own policy while matching a stronger teacher. This addresses the prefix mismatch of offline distillation, but early student rollouts can still be poor, placing teacher supervision on weak or low-quality prefixes. We propose Trust-Region behavior Blending (TRB), a warmup method that replaces the early rollout policy with the closest-to-teacher behavior policy inside a student-centered KL trust region, while keeping the per-prefix reverse-KL OPD loss unchanged. The KL budget is annealed to zero, so training returns to pure student rollouts after warmup. Across two math-reasoning distillation settings, TRB attains the strongest average among the compared methods.

Language models fine-tuned with reinforcement learning typically optimize for task reward, ignoring multi-agent strategic structure. Because these agents condition on natural language game-state descriptions and emit actions through free-form generation, strategic failure modes -- exploiting weaker opponents, coordinating on harmful equilibria, and externalizing costs are inseparable from the language interface itself. We propose Safe Equilibrium Policy Optimization (\sepo{}), a training objective that augments expected payoff with explicit penalties for exploitability, collusion risk, and externality cost. We implement \sepo{} as a reward signal for Group Relative Policy Optimization (GRPO), applied to Gemma4 E4B-it and Qwen3.5-4B after supervised fine-tuning (SFT). Evaluated across five strategic domains: Iterated Prisoner's Dilemma, repeated auctions, two negotiation variants, and Kuhn Poker. \sepo{} achieves zero exploit-pool advantage in Kuhn Poker for both models, outperforms the base model on safety in four domains, and corrects the over-cooperative behavior introduced by SFT. In negotiation, \sepo{} achieves a positive-safety outcome and only the positive normalized relative advantage of any negotiation configuration. Ablation experiments confirm that per-rollout exploit computation is necessary: a shared constant penalty cancels in GRPO advantage normalization (constant control-variate property), producing zero gradient. To support further research in strategic safety for agents, we release our code and SFT datasets.

In reinforcement learning (RL), agents benefit from exploration only because they repeatedly encounter similar states: trying different actions can improve performance or reduce uncertainty; without such retries, a greedy policy is optimal. We formalize this intuition with ReMax, an objective that evaluates a policy by the expected maximum return over $M$ samples, where $M$ is a positive integer, while accounting for return uncertainty. Optimizing this objective induces stochastic exploration as an emergent property, without explicit bonus terms. For efficient policy optimization, we derive a new policy-gradient formulation for ReMax and introduce ReMax PPO (RePPO), a PPO variant that optimizes ReMax while generalizing the discrete retry count $M$ to a continuous parameter $m > 0$, enabling fine-grained control of exploration. Empirically, RePPO promotes exploration, without any explicit exploration bonuses, on the MinAtar and Craftax benchmarks.

Maximum entropy reinforcement learning (MaxEnt-RL) enables robust exploration, yet practical implementations often restrict policies to simple Gaussians. While recent approaches incorporate expressive generative policies via importance-weighted supervised learning, they are prone to importance weight collapse, which limits their scalability in high-dimensional action spaces. Our key insight is to mitigate this limitation by localizing the sampling region, avoiding the weight degeneracy induced by importance sampling over the entire action space. To instantiate this insight, we introduce \textbf{FLAG} (\textbf{F}low policy with \textbf{L}atent-\textbf{A}ugmented \textbf{G}uidance). FLAG augments the state space with a flow latent variable and optimizes a provably consistent proxy MaxEnt-RL objective. We empirically demonstrate that FLAG enables expressive policy optimization with limited importance samples and scales to high-dimensional control tasks. Furthermore, FLAG achieves state-of-the-art performance across challenging benchmarks. Our project webpage: https://flag-rl.github.io/

We study when large language models (LLMs) can serve as effective black-box policy optimizers for reinforcement learning (RL) tasks, i.e., when can we replace classical RL algorithms with an LLM? We explore this question by introducing Prompted Policy Optimization (PromptPO), an iterative method that prompts an LLM with Python descriptions of the state space, action space, and reward function, then has it generate and refine executable policies based on rollout feedback. Across hard exploration environments, Meta-World robotics tasks, and several real-world control problems, PromptPO often matches or exceeds the performance of standard RL baselines while using substantially fewer environment interactions. To maximize expected return, and without further explicit prompting, the policies PromptPO outputs range from tuned proportional controllers or rule-based plans to policies that run planning algorithms like value iteration. Our results demonstrate that LLM-based policy optimization is sufficient when the LLM can leverage prior knowledge about the environment or optimization strategy. PromptPO underperforms standard RL baselines in MuJoCo domains. This demonstrates possible limitations of LLM-based policy optimization to settings that requiring fine-grained continuous control.

We investigate a narrow but common failure mode of GRPO-style reinforcement learning in the context of sparse verifiable rewards: early updates contain more responses with negative advantages than those with positive advantages, while response-level length normalization ties the magnitude of the update to the length of the output. We propose Hysteretic Policy Optimization (HPO), a minimal modification of GRPO that reduces the weight of negative-advantage updates and replaces per-response length normalization with mean-length normalization. We further introduce Adaptive HPO (A-HPO), which sets the hysteretic weight based on batch-level advantage-sign statistics, thereby removing the need for tuning a fixed hysteretic weight. In our TeleLogs and Countdown experiments, A-HPO improves the reward per update compared to GRPO, with the largest gains in early sparse reward regimes. On TeleLogs, A-HPO achieves a final reward of 0.84, outperforming SAPO by 5%, GSPO by 11%, and GRPO by 15%, while maintaining a comparable response-length. On Countdown, A-HPO achieves the largest gains in initial and most difficult configurations across 1.5B-7B models. Ablation studies on the hysteretic weight show that the gains of A-HPO come from better balancing the contributions of positive and negative advantages compared to positive-only or fully symmetric updates.

When a large language model under reinforcement learning commits a wrong reasoning step early in a trajectory, standard algorithms force it to keep generating until the maximum horizon, spending compute on tokens that never receive positive reward and polluting advantage estimates with post-failure noise. We propose ESPO (Early-Stopping Proximal Policy Optimization), which detects trajectory failure on-the-fly and terminates rollouts early. At each generation step, ESPO computes a surrogate regret using only the logits already computed during sampling, and terminates when the smoothed cumulative regret significantly exceeds its estimated values. Truncated trajectories are treated as absorbing failure states with a terminal reward, concentrating negative temporal-difference (TD) errors near the detected failure step without any additional reward model or human annotation. On DeepSeek-R1-Distill-Qwen-7B trained for mathematical reasoning, ESPO surpasses PPO on AIME2024 (46.28% vs. 45.25%), AMC2023 (85.83% vs. 82.94%), and MATH-500 (87.42% vs. 85.43%), while saving more than 20% rollout tokens cumulatively.

Reinforcement learning with multinomial logistic (MNL) function approximation has become an important framework due to its flexibility and broad applicability. While existing studies have established regret guarantees under worst-case analysis, they do not capture how performance depends on the variability of the interaction between the learner and the environment. In this paper, we develop a new theoretical analysis for MNL-based Markov decision processes that yields explicit variance-adaptive regret bounds. Our algorithm is computationally efficient and achieves the instance-wise optimal rate of regret, narrowing the gap between upper and lower bounds. Our numerical experiments validate that our method learns optimal policies more efficiently than conventional approaches.

Reinforcement learning algorithms are commonly analyzed (and designed) under the Markov assumption. This is unrealistic, as most environments encountered in practice are either partially observable, or require function approximation that restricts the agent to access non-Markovian state features. We consider the problem of learning an optimal reactive policy in a finite environment with deterministic observations (or equivalently, hard state aggregation). We introduce a new algorithm, Committed Q-learning, and prove almost-sure convergence to the optimal reactive policy under an intuitive assumption we call rewire-robustness. This assumption is strictly weaker than the $q_\star$-realizability condition used in prior work. Our algorithm is a variant of classical Q-learning in which the behavior policy commits to a single action upon entering a feature, and only resamples actions when the observed feature changes. A crucial part of our analysis is the introduction of quasi-Markov environments.

Large scale reinforcement learning has become a central tool for improving reasoning in large language models. At this scale, generation is often lagged or asynchronous, so updates are performed on data collected by older policies. This makes learning inherently off-policy. Most existing approaches nevertheless remain rooted in PPO-style trust-region objectives, treating training as approximately on-policy and using importance weights to correct distribution mismatch. These corrections can introduce high variance, destabilize optimization, and accelerate entropy collapse. Recent work suggests an alternative: rather than correcting the mismatch, one can embrace off-policy data and remove importance weights, often yielding stronger algorithms. In this paper, we provide an intuitive construction of off-policy objectives that include successful off-policy objectives and show that their effectiveness can be understood through implicit pessimism: they optimize toward target policies that are more conservative than their nominal objectives suggest. This perspective explains why some particular implementation choices improve stability: they implicitly control the effective target distribution. We then propose a principled modification that stabilize this induced distribution and improve off-policy learning.

We address the problem of learning to assign prediction tasks to one agent from a set of available agents, including human decision-makers and AI models. We focus on sequential learning of agent expertise and assignment policies where each agent is constrained to handle a fraction of tasks. We provide a general theoretical characterization of this problem in terms of agent capacities, differences in agent expertise, and task context. We then develop a framework of sequential explore-exploit policy-learning algorithms that seek to maximize overall performance. Experimental results over a variety of tabular, image, and text prediction tasks demonstrate systematic gains from our policy-learning algorithms relative to non-contextual baselines across different types of agents, including LLMs and humans.

Offline policy improvement faces an inherent conflict between maximizing value and fitting the data distribution. While in-sample weighted regression is stable, it suffers from over-conservatism that suppresses high-value actions in the distribution tail; conversely, gradient-based approaches often exhibit a fitting-optimization conflict of gradients, which drives the policy off the data manifold. To address this, we propose Support-Preserving Action Rectification (SPAR), which reframes global learning as a local residual rectification anchored to a frozen pure behavior cloning policy. This framework performs fine-grained fitting and local policy improvement in the residual space, thereby contracting the search space. We further introduce Latent Self-Imitation, utilizing a latent-sampling weighted-regression mechanism to address fitting-improvement gradient conflict in the residual space. Theoretically, we prove this mechanism eliminates the manifold-normal drift of standard value gradients, while extensive D4RL experiments show SPAR extracts significant gains from suboptimal baselines to achieve state-of-the-art performance.

Reinforcement learning with verifiable rewards has become a standard recipe for improving the reasoning abilities of large language models. Existing algorithms face a tradeoff between computational efficiency and sample efficiency in value estimation and policy learning. We introduce BASIS, a critic-free post-training algorithm designed to address this tradeoff. At each online training step, BASIS samples only one rollout per prompt, but leverages rich information across prompts in the entire batch to improve value function estimation. Our experiments demonstrate that BASIS reduces MSE in value function estimation by 69% compared to REINFORCE++, a representative single-rollout baseline, and achieves lower MSE with one rollout than group mean estimators with 8 rollouts. This improvement in value estimation translates to better policy optimization: using substantially less training time, BASIS achieves performance close to multi-rollout GRPO-type baselines and often outperforms single-rollout REINFORCE-type baselines.