Offline RL (ORL) promises safe and sample-efficient deployment but existing methods rely on undocumented online interactions for hyperparameter tuning and lack reliable fully offline estimates of initial online performance. We introduce SOReL, a fully offline Bayesian model-based RL method that learns a posterior over dynamics, estimates policy value via predictive uncertainty, and enables complete offline hyperparameter selection. We further propose TOReL, which extends this tuning framework to arbitrary model-free and model-based ORL algorithms. We provide a regret analysis showing that Bayesian offline RL achieves the minimax-optimal parametric rate under standard regularity conditions. Together, our methods establish a practical and theoretically grounded framework for fully offline RL.
Scaling reinforcement learning (RL) workloads often requires distributing environment simulation across compute clusters. Existing frameworks entangle simulation, learning logic, and orchestration into monolithic systems, limiting modularity and reusability. We present ClusterEnv, a lightweight, learner-agnostic interface for distributed environment execution that mirrors the Gymnasium API. ClusterEnv introduces the DETACH pattern, which decouples simulation from training by offloading reset() and step() operations to remote workers while keeping learning centralized. To address policy staleness in distributed execution, we propose Adaptive Actor Policy Synchronization (AAPS), a divergence-triggered update mechanism that reduces synchronization overhead without sacrificing performance. ClusterEnv integrates cleanly into existing RL pipelines, supports both on-policy and off-policy methods, and requires minimal code changes. Experiments on discrete control tasks demonstrate that AAPS achieves high sample efficiency with significantly fewer weight updates. Source code is available at https://github.com/rodlaf/ClusterEnv.
Tractography algorithms leverage diffusion MRI to reconstruct the fibrous architecture of the brain's white matter. Among machine learning approaches, reinforcement learning (RL) has emerged as a promising framework for tractography, outperforming traditional methods in several key aspects. TractOracle-RL, a recent RL-based approach, reduces false positives by incorporating anatomical priors into the training process via a reward-based mechanism. In this paper, we investigate four extensions of the original TractOracle-RL framework by integrating recent advances in RL, and we evaluate their performance across five diverse diffusion MRI datasets. Results demonstrate that combining an oracle with the RL framework consistently leads to robust and reliable tractography, regardless of the specific method or dataset used. We also introduce a novel RL training scheme called Iterative Reward Training (IRT), inspired by the Reinforcement Learning from Human Feedback (RLHF) paradigm. Instead of relying on human input, IRT leverages bundle filtering methods to iteratively refine the oracle's guidance throughout training. Experimental results show that RL methods trained with oracle feedback significantly outperform widely used tractography techniques in terms of accuracy and anatomical validity.
Large language models (LLMs) excel at logical and algorithmic reasoning, yet their emotional intelligence (EQ) still lags far behind their cognitive prowess. While reinforcement learning from verifiable rewards (RLVR) has advanced in other domains, its application to dialogue-especially for emotional intelligence-remains underexplored. In this work, we introduce RLVER, the first end-to-end reinforcement learning framework that leverages verifiable emotion rewards from simulated users to cultivate higher-order empathetic abilities in LLMs. Within this framework, self-consistent affective simulated users engage in dialogue rollouts and produce deterministic emotion scores during conversations, serving as reward signals to guide the LLM's learning. Fine-tuning publicly available Qwen2.5-7B-Instruct model with PPO boosts its Sentient-Benchmark score from 13.3 to 79.2 while largely preserving mathematical and coding competence. Extensive experiments reveal that: (i) RLVER consistently improves multiple dialogue capabilities; (ii) Thinking and non-thinking models show distinct trends--thinking models excel in empathy and insight, while non-thinking models favor action; (iii) GRPO often yields stable gains, while PPO can push certain capabilities to a higher ceiling; (iv) More challenging environments are not always better-moderate ones can yield stronger outcomes. Our results show that RLVER is a practical route toward emotionally intelligent and broadly capable language agents.
We present Step-wise Policy for Rare-tool Knowledge (SPaRK), a novel reinforcement learning framework that teaches large language models to explore diverse tool usage patterns beyond conventional high-temperature sampling. Building on recent advances in step-wise reinforcement learning, we introduce a dual-objective reward system that simultaneously optimizes for answer quality and tool diversity, training a Llama-3.1 8B model through offline PPO on synthetically generated trajectories from the MMLU-Pro dataset. Our approach uniquely employs a rarity-first exploitation strategy where a GPT-4o judge scores candidate actions across eight distinct tools plus chain-of-thought reasoning, with the policy favoring less-frequently used but still viable tools to encourage systematic exploration. Empirical results demonstrate that SPaRK achieves competitive performance across 14 MMLU-Pro categories while exhibiting significantly higher entropy in tool selection compared to both baseline and supervised fine-tuning approaches, suggesting that algorithmic exploration through explicit tool diversity can enhance reasoning capabilities without sacrificing accuracy.
Achieving fast and stable off-policy learning in deep reinforcement learning (RL) is challenging. Most existing methods rely on semi-gradient temporal-difference (TD) methods for their simplicity and efficiency, but are consequently susceptible to divergence. While more principled approaches like Gradient TD (GTD) methods have strong convergence guarantees, they have rarely been used in deep RL. Recent work introduced the Generalized Projected Bellman Error ($\GPBE$), enabling GTD methods to work efficiently with nonlinear function approximation. However, this work is only limited to one-step methods, which are slow at credit assignment and require a large number of samples. In this paper, we extend the $\GPBE$ objective to support multistep credit assignment based on the $\lambda$-return and derive three gradient-based methods that optimize this new objective. We provide both a forward-view formulation compatible with experience replay and a backward-view formulation compatible with streaming algorithms. Finally, we evaluate the proposed algorithms and show that they outperform both PPO and StreamQ in MuJoCo and MinAtar environments, respectively. Code available at https://github.com/esraaelelimy/gtd\_algos
Kernel development in deep learning requires optimizing computational units across hardware while balancing memory management, parallelism, and hardware-specific optimizations through extensive empirical tuning. Although domain-specific languages like Triton simplify GPU programming by abstracting low-level details, developers must still manually tune critical parameters such as tile sizes and memory access patterns through iterative experimentation, creating substantial barriers to optimal performance and wider adoption. In this work, we introduce AutoTriton, the first model dedicated to Triton programming powered by reinforcement learning (RL). AutoTriton performs supervised fine-tuning (SFT) to be equipped with essential Triton programming expertise using a high-quality data gathering pipeline, and conducts RL with Group Relative Policy Optimization (GRPO) algorithm, combining a rule-based reward and an execution-based reward to further improve Triton programming ability, sequentially. Experiments across five evaluation channels of TritonBench and KernelBench illustrate that our 8B model AutoTriton achieves performance comparable to mainstream large models, including Claude-4-Sonnet and DeepSeek-R1-0528. Further experimental analysis demonstrates the crucial role of each module within AutoTriton, including the SFT stage, the RL stage, and the reward design strategy. These findings underscore the promise of RL for automatically generating high-performance kernels, and since high-performance kernels are core components of AI systems, this breakthrough establishes an important foundation for building more efficient AI systems. The model and code will be available at https://github.com/AI9Stars/AutoTriton.
Translating natural language mathematical statements into formal, executable code is a fundamental challenge in automated theorem proving. While prior work has focused on generation and compilation success, little attention has been paid to the critic phase-the evaluation of whether generated formalizations truly capture the semantic intent of the original problem. In this paper, we introduce CriticLean, a novel critic-guided reinforcement learning framework that elevates the role of the critic from a passive validator to an active learning component. Specifically, first, we propose the CriticLeanGPT, trained via supervised fine-tuning and reinforcement learning, to rigorously assess the semantic fidelity of Lean 4 formalizations. Then, we introduce CriticLeanBench, a benchmark designed to measure models' ability to distinguish semantically correct from incorrect formalizations, and demonstrate that our trained CriticLeanGPT models can significantly outperform strong open- and closed-source baselines. Building on the CriticLean framework, we construct FineLeanCorpus, a dataset comprising over 285K problems that exhibits rich domain diversity, broad difficulty coverage, and high correctness based on human evaluation. Overall, our findings highlight that optimizing the critic phase is essential for producing reliable formalizations, and we hope our CriticLean will provide valuable insights for future advances in formal mathematical reasoning.
This paper introduces TransDreamerV3, a reinforcement learning model that enhances the DreamerV3 architecture by integrating a transformer encoder. The model is designed to improve memory and decision-making capabilities in complex environments. We conducted experiments on Atari-Boxing, Atari-Freeway, Atari-Pong, and Crafter tasks, where TransDreamerV3 demonstrated improved performance over DreamerV3, particularly in the Atari-Freeway and Crafter tasks. While issues in the Minecraft task and limited training across all tasks were noted, TransDreamerV3 displays advancement in world model-based reinforcement learning, leveraging transformer architectures.
Transfer learning in Reinforcement Learning (RL) enables agents to leverage knowledge from source tasks to accelerate learning in target tasks. While prior work, such as the Attend, Adapt, and Transfer (A2T) framework, addresses negative transfer and selective transfer, other critical challenges remain underexplored. This paper introduces the Generalized Adaptive Transfer Network (GATN), a deep RL architecture designed to tackle task generalization across domains, robustness to environmental changes, and computational efficiency in transfer. GATN employs a domain-agnostic representation module, a robustness-aware policy adapter, and an efficient transfer scheduler to achieve these goals. We evaluate GATN on diverse benchmarks, including Atari 2600, MuJoCo, and a custom chatbot dialogue environment, demonstrating superior performance in cross-domain generalization, resilience to dynamic environments, and reduced computational overhead compared to baselines. Our findings suggest GATN is a versatile framework for real-world RL applications, such as adaptive chatbots and robotic control.
Training large language models (LLMs) to act as autonomous agents for multi-turn, long-horizon tasks remains significant challenges in scalability and training efficiency. To address this, we introduce L-Zero (L0), a scalable, end-to-end training pipeline for general-purpose agents. Featuring a low-cost, extensible, and sandboxed concurrent agent worker pool, L0 lowers the barrier for applying reinforcement learning in complex environments. We also introduce NB-Agent, the agent scaffold within L0, which operates in a "code-as-action" fashion via a Read-Eval-Print-Loop (REPL). We evaluate L0 on factuality question-answering benchmarks. Our experiments demonstrate that a base model can develop robust problem-solving skills using solely Reinforcement Learning with Verifiable Rewards (RLVR). On the Qwen2.5-7B-Instruct model, our method boosts accuracy on SimpleQA from 30 % to 80 % and on HotpotQA from 22 % to 41 %. We have open-sourced the entire L0 system, including our L0 series models, the NB-Agent, a complete training pipeline, and the corresponding training recipes on (https://github.com/cmriat/l0).
Deep Reinforcement Learning (DRL) has achieved remarkable success in complex sequential decision-making tasks, such as playing Atari 2600 games and mastering board games. A critical yet underexplored aspect of DRL is the temporal scale of action execution. We propose a novel paradigm that integrates contextual bandits with DRL to adaptively select action durations, enhancing policy flexibility and computational efficiency. Our approach augments a Deep Q-Network (DQN) with a contextual bandit module that learns to choose optimal action repetition rates based on state contexts. Experiments on Atari 2600 games demonstrate significant performance improvements over static duration baselines, highlighting the efficacy of adaptive temporal abstractions in DRL. This paradigm offers a scalable solution for real-time applications like gaming and robotics, where dynamic action durations are critical.
While large-scale unsupervised language models (LMs) capture broad world knowledge and reasoning capabilities, steering their behavior toward desired objectives remains challenging due to the lack of explicit supervision. Existing alignment techniques, such as reinforcement learning from human feedback (RLHF), rely on training a reward model and performing reinforcement learning to align with human preferences. However, RLHF is often computationally intensive, unstable, and sensitive to hyperparameters. To address these limitations, Direct Preference Optimization (DPO) was introduced as a lightweight and stable alternative, enabling direct alignment of language models with pairwise preference data via classification loss. However, DPO and its extensions generally assume a single static preference distribution, limiting flexibility in multi-objective or dynamic alignment settings. In this paper, we propose a novel framework: Multi-Preference Lambda-weighted Listwise DPO, which extends DPO to incorporate multiple human preference dimensions (e.g., helpfulness, harmlessness, informativeness) and enables dynamic interpolation through a controllable simplex-weighted formulation. Our method supports both listwise preference feedback and flexible alignment across varying user intents without re-training. Empirical and theoretical analysis demonstrates that our method is as effective as traditional DPO on static objectives while offering greater generality and adaptability for real-world deployment.
This paper introduces an efficient Residual Reinforcement Learning (RRL) framework for voltage control in active distribution grids. Voltage control remains a critical challenge in distribution grids, where conventional Reinforcement Learning (RL) methods often suffer from slow training convergence and inefficient exploration. To overcome these challenges, the proposed RRL approach learns a residual policy on top of a modified Sequential Droop Control (SDC) mechanism, ensuring faster convergence. Additionally, the framework introduces a Local Shared Linear (LSL) architecture for the Q-network and a Transformer-Encoder actor network, which collectively enhance overall performance. Unlike several existing approaches, the proposed method relies solely on inverters' measurements without requiring full state information of the power grid, rendering it more practical for real-world deployment. Simulation results validate the effectiveness of the RRL framework in achieving rapid convergence, minimizing active power curtailment, and ensuring reliable voltage regulation.
Large Language Models (LLMs), particularly slow-thinking models, often exhibit severe hallucination, outputting incorrect content due to an inability to accurately recognize knowledge boundaries during reasoning. While Reinforcement Learning (RL) can enhance complex reasoning abilities, its outcome-oriented reward mechanism often lacks factual supervision over the thinking process, further exacerbating the hallucination problem. To address the high hallucination in slow-thinking models, we propose Knowledge-enhanced RL, KnowRL. KnowRL guides models to perform fact-based slow thinking by integrating a factuality reward, based on knowledge verification, into the RL training process, helping them recognize their knowledge boundaries. KnowRL guides models to perform fact-based slow thinking by integrating a factuality reward, based on knowledge verification, into the RL training process, helping them recognize their knowledge boundaries. This targeted factual input during RL training enables the model to learn and internalize fact-based reasoning strategies. By directly rewarding adherence to facts within the reasoning steps, KnowRL fosters a more reliable thinking process. Experimental results on three hallucination evaluation datasets and two reasoning evaluation datasets demonstrate that KnowRL effectively mitigates hallucinations in slow-thinking models while maintaining their original strong reasoning capabilities. Our code is available at https://github.com/zjunlp/KnowRL.
In this paper, we investigate the use of small datasets in the context of offline reinforcement learning (RL). While many common offline RL benchmarks employ datasets with over a million data points, many offline RL applications rely on considerably smaller datasets. We show that offline RL algorithms can overfit on small datasets, resulting in poor performance. To address this challenge, we introduce "Sparse-Reg": a regularization technique based on sparsity to mitigate overfitting in offline reinforcement learning, enabling effective learning in limited data settings and outperforming state-of-the-art baselines in continuous control.
In reinforcement learning (RL), value-based algorithms learn to associate each observation with the states and rewards that are likely to be reached from it. We observe that many self-supervised image pre-training methods bear similarity to this formulation: learning features that associate crops of images with those of nearby views, e.g., by taking a different crop or color augmentation. In this paper, we complete this analogy and explore a method that directly casts pre-training on unlabeled image data like web crawls and video frames as an RL problem. We train a general value function in a dynamical system where an agent transforms an image by changing the view or adding image augmentations. Learning in this way resembles crop-consistency self-supervision, but through the reward function, offers a simple lever to shape feature learning using curated images or weakly labeled captions when they exist. Our experiments demonstrate improved representations when training on unlabeled images in the wild, including video data like EpicKitchens, scene data like COCO, and web-crawl data like CC12M.
Reinforcement learning (RL) with tree search has demonstrated superior performance in traditional reasoning tasks. Compared to conventional independent chain sampling strategies with outcome supervision, tree search enables better exploration of the reasoning space and provides dense, on-policy process rewards during RL training but remains under-explored in On-Policy LLM RL. We propose TreeRL, a reinforcement learning framework that directly incorporates on-policy tree search for RL training. Our approach includes intermediate supervision and eliminates the need for a separate reward model training. Existing approaches typically train a separate process reward model, which can suffer from distribution mismatch and reward hacking. We also introduce a cost-effective tree search approach that achieves higher search efficiency under the same generation token budget by strategically branching from high-uncertainty intermediate steps rather than using random branching. Experiments on challenging math and code reasoning benchmarks demonstrate that TreeRL achieves superior performance compared to traditional ChainRL, highlighting the potential of tree search for LLM. TreeRL is open-sourced at https://github.com/THUDM/TreeRL.
Different base language model families, such as Llama and Qwen, exhibit divergent behaviors during post-training with reinforcement learning (RL), especially on reasoning-intensive tasks. What makes a base language model suitable for reinforcement learning? Gaining deeper insight into this question is essential for developing RL-scalable foundation models of the next generation. In this work, we investigate how mid-training strategies shape RL dynamics, focusing on two representative model families: Qwen and Llama. Our study reveals that (1) high-quality mathematical corpora, such as MegaMath-Web-Pro, significantly improve both base model and RL performance, while existing alternatives (e.g., FineMath-4plus) fail to do so; (2) further adding QA-style data, particularly long chain-of-thought (CoT) reasoning examples, enhances RL outcomes, and instruction data further unlocks this effect; (3) while long-CoT improves reasoning depth, it can also induce verbosity of model responses and unstability of RL training, underscoring the importance of data formatting; (4) scaling mid-training consistently leads to stronger downstream RL performance. Building on these insights, we introduce a two-stage mid-training strategy, Stable-then-Decay, in which base models are first trained on 200B tokens with a constant learning rate, followed by 20B tokens across three CoT-focused branches with learning rate decay. This yields OctoThinker, a family of models demonstrating strong RL compatibility and closing the performance gap with more RL-friendly model families, i.e., Qwen. We hope our work will help shape pre-training strategies for foundation models in the RL era. To support further research, we release our open-source models along with a curated math reasoning-intensive corpus of over 70 billion tokens (i.e., MegaMath-Web-Pro-Max).
Large Language Models (LLMs) exhibit remarkable code generation capabilities but falter when adapting to frequent updates in external library APIs. This critical limitation, stemming from reliance on outdated API knowledge from their training data, even with access to current documentation, impedes reliable code generation in dynamic environments. To tackle this issue, we propose ReCode (rule-based Reinforcement learning for Code Update), a novel framework that mimics human programmer adaptation to API changes. Specifically, we construct a dataset of approximately 2,000 data entries to train the LLMs to perform version migration based on updated information. Then, we introduce a modified string similarity metric for code evaluation as the reward for reinforcement learning. Our experiments demonstrate that ReCode substantially boosts LLMs' code generation performance in dynamic API scenarios, especially on the unseen CodeUpdateArena task. Crucially, compared to supervised fine-tuning, ReCode has less impact on LLMs' general code generation abilities. We apply ReCode on various LLMs and reinforcement learning algorithms (GRPO and DAPO), all achieving consistent improvements. Notably, after training, Qwen2.5-Coder-7B outperforms that of the 32B parameter code instruction-tuned model and the reasoning model with the same architecture. Code is available at https://github.com/zjunlp/ReCode.