One of the primary drivers of the success of machine learning methods in open-world perception settings, such ascomputer vision [19]and NLP [8],has been the ability ofhigh-capacity function approximators, suchasdeepneuralnetworks,tolearngeneralizable modelsfromlargeamountsof data.

Letπ0( |s)beaGaussianbehavioral reference policy with meanµ0(s) and variance σ20(s), and let π( |s) be an online policy with reparameterization at = fφ( t;st)andrandomvector t. Whilstentropyregularization partially mitigates the collapse of predictive variance away from the expert demonstrations, we still observe the wrong trend similar to Figure 1 with predictive variances high near the expert demonstrations andlowonunseen data. AWAC performs online fine-tuning of a policy pre-trained on offline. Themethod requires additional off-policy data to be generated to saturate the replay buffer, thereby requiring ahidden number ofenvironment interactions that donotinvolvelearning. To mitigate this, in practice, BRAC adds an entropy bonus to the supervised learning objective which stabilizes the variance around the training set but has no guarantees away from thedata.

In deep reinforcement learning, policy optimization methods need to deal with issues such as function approximation and the reuse of off-policy data. Standard policy gradient methods do not handle off-policy data well, leading to premature convergence and instability. This paper introduces a method to stabilize policy optimization when off-policy data are reused. The idea is to include a Bregman divergence between the behavior policy that generates the data and the current policy to ensure small and safe policy updates with off-policy data. The Bregman divergence is calculated between the state distributions of two policies, instead of only on the action probabilities, leading to a divergence augmentation formulation. Empirical experiments on Atari games show that in the data-scarce scenario where the reuse of off-policy data becomes necessary, our method can achieve better performance than other state-of-the-art deep reinforcement learning algorithms.

Evaluating policies using off-policy data is crucial for applying reinforcement learning to real-world problems such as healthcare and autonomous driving. Previous methods for (OPE) generally suffer from high variance or irreducible bias, leading to unacceptably high prediction errors. In this work, we introduce STAR, a framework for OPE that encompasses a broad range of estimators -- which include existing OPE methods as special cases -- that achieve lower mean squared prediction errors. STAR leverages state abstraction to distill complex, potentially continuous problems into compact, discrete models which we call (ARPs). Predictions from ARPs estimated from off-policy data are provably consistent (asymptotically correct). Rather than proposing a specific estimator, we present a new framework for OPE and empirically demonstrate that estimators within STAR outperform existing methods. The best STAR estimator outperforms baselines in all twelve cases studied, and even the median STAR estimator surpasses the baselines in seven out of the twelve cases.

Offline reinforcement learning, wherein one uses off-policy data logged by a fixed behavior policy to evaluate and learn new policies, is crucial in applications where experimentation is limited such as medicine. We study the estimation of policy value and gradient of a deterministic policy from off-policy data when actions are continuous. Targeting deterministic policies, for which action is a deterministic function of state, is crucial since optimal policies are always deterministic (up to ties). In this setting, standard importance sampling and doubly robust estimators for policy value and gradient fail because the density ratio does not exist. To circumvent this issue, we propose several new doubly robust estimators based on different kernelization approaches. We analyze the asymptotic mean-squared error of each of these under mild rate conditions for nuisance estimators. Specifically, we demonstrate how to obtain a rate that is independent of the horizon length.

We propose an estimator and confidence interval for computing the value of a policy from off-policy data in the contextual bandit setting. To this end we apply empirical likelihood techniques to formulate our estimator and confidence interval as simple convex optimization problems. Using the lower bound of our confidence interval, we then propose an off-policy policy optimization algorithm that searches for policies with large reward lower bound. We empirically find that both our estimator and confidence interval improve over previous proposals in finite sample regimes. Finally, the policy optimization algorithm we propose outperforms a strong baseline system for learning from off-policy data.

Reinforcement learning (RL) is a critical component of large language model (LLM) post-training. However, on-policy algorithms used for post-training are not naturally robust to a diversified content of experience replay buffers, which asynchronous off-policy actors can efficiently populate in parallel to training. We propose efficiently learning on such off-policy data via Trajectory Balance with Asynchrony (TBA), an approach to asynchronous RL for LLMs that leverages the principled off-policy TB objective. On math, preference-tuning, and automated red-teaming tasks, we post-train models ranging from Pythia 410M to Qwen 2.5 7B, finding TBA offers speed and performance boosts over strong baselines like Online DPO and Dr. GRPO. Beyond TBA's performance benefits (high accuracy even as asynchrony grows) and speedups ($4\times$ or more), we show its reward- and recency-prioritizing sampling enable further gains as data generation is scaled. Our code is available at https://github.com/bbartoldson/TBA.

Reinforcement Learning (RL) has become a key driver for enhancing the long chain-of-thought (CoT) reasoning capabilities of Large Language Models (LLMs). However, prevalent methods like GRPO often fail when task difficulty exceeds the model's capacity, leading to reward sparsity and inefficient training. While prior work attempts to mitigate this using off-policy data, such as mixing RL with Supervised Fine-Tuning (SFT) or using hints, they often misguide policy updates In this work, we identify a core issue underlying these failures, which we term low training affinity. This condition arises from a large distributional mismatch between external guidance and the model's policy. To diagnose this, we introduce Affinity, the first quantitative metric for monitoring exploration efficiency and training stability. To improve Affinity, we propose HINT: Helping Ineffective rollouts Navigate Towards effectiveness, an adaptive hinting framework. Instead of providing direct answers, HINT supplies heuristic hints that guide the model to discover solutions on its own, preserving its autonomous reasoning capabilities. Extensive experiments on mathematical reasoning tasks show that HINT consistently outperforms existing methods, achieving state-of-the-art results with models of various scales, while also demonstrating significantly more stable learning and greater data efficiency.Code is available on Github.

Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) are two prominent post-training paradigms for refining the capabilities and aligning the behavior of Large Language Models (LLMs). Existing approaches that integrate SFT and RL often face the risk of disrupting established response patterns and inducing overfitting to expert data. To address this, we present a novel investigation into the unified view of SFT and RL through an off-policy versus on-policy lens. We propose CHORD, a framework for Controllable Harmonization of On- and Off-Policy Reinforcement Learning via Dynamic Weighting, which reframes SFT not as a separate stage but as a dynamically weighted auxiliary objective within the on-policy RL process. Based on an analysis of off-policy expert data's influence at both holistic and granular levels, we incorporate a dual-control mechanism in CHORD. Specifically, the framework first employs a global coefficient to holistically guide the transition from off-policy imitation to on-policy exploration, and then applies a token-wise weighting function that enables granular learning from the expert, which promotes on-policy exploration and mitigates disruption from off-policy data. We conduct extensive experiments on mathematical reasoning problems and practical tool-use tasks, providing empirical evidence that CHORD achieves a stable and efficient learning process. By effectively harmonizing off-policy expert data with on-policy exploration, CHORD demonstrates significant improvements over baselines. We release the implementation at https://github.com/modelscope/Trinity-RFT/tree/main/examples/mix_chord to inspire further research.