论文检索

Reinforcement Learning · Deep RL

Chongyi Zheng, Mahsa Bastankhah, Grace Liu, Benjamin Eysenbach

Goal-conditioned reinforcement learning learns to reach goals instead of optimizing hand-crafted rewards. Despite its popularity, the community often categorizes goal-conditioned reinforcement learning as a special case of reinforcement learning. In this post, we aim to build a direct conversion from any reward-maximization reinforcement learning problem to a goal-conditioned reinforcement learning problem, and to draw connections with the stochastic shortest path framework. Our conversion provides a new perspective on the reinforcement learning problem: *maximizing rewards is equivalent to reaching some goals*.

Reinforcement Learning · Deep RL

Chongyi Zheng, Seohong Park, Sergey Levine, Benjamin Eysenbach

Large-scale pre-training has fundamentally changed how machine learning research is done today: large foundation models are trained once, and then can be used by anyone in the community (including those without data or compute resources to train a model from scratch) to adapt and fine-tune to specific tasks. Applying this same framework to reinforcement learning (RL) is appealing because it offers compelling avenues for addressing core challenges in RL, including sample efficiency and robustness. However, there remains a fundamental challenge to pre-train large models in the context of RL: actions have long-term dependencies, so training a foundation model that reasons across *time* is important. Recent advances in generative AI have provided new tools for modeling highly complex distributions. In this paper, we build a probabilistic model to predict which states an agent will visit in the temporally distant future (i.e., an occupancy measure) using flow matching. As large datasets are often constructed by many distinct users performing distinct tasks, we include in our model a latent variable capturing the user intention. This intention increases the expressivity of our model, and enables adaptation with generalized policy improvement. We call our proposed method **intention-conditioned flow occupancy models (InFOM)**. Comparing with alternative methods for pre-training, our experiments on $36$ state-based and $4$ image-based benchmark tasks demonstrate that the proposed method achieves $1.8 \times$ median improvement in returns and increases success rates by $36\\%$.

Reinforcement Learning · Deep RL

Yuan Pu, Yazhe Niu, Jia Tang, Junyu Xiong, Shuai Hu, Hongsheng Li

In heterogeneous multi-task decision-making, tasks not only exhibit diverse observation and action spaces but also vary substantially in their underlying complexities. While conventional multi-task world models like UniZero excel in single-task settings, we find that when handling a broad and diverse suite of tasks, gradient conflicts and the loss of model plasticity often constrain their sample efficiency. In this work, we address these challenges from two complementary perspectives: the single learning iteration and the overall learning process. First, to mitigate the gradient conflicts, we systematically investigate key architectural designs for extending UniZero. Our investigation identifies a Mixture-of-Experts (MoE) architecture as the most effective approach. We demonstrate, both theoretically and empirically, that this architecture alleviates gradient conflicts by routing task-specific representations to specialized sub-networks. This finding leads to our proposed model, \textit{ScaleZero}. Second, to dynamically allocate model capacity throughout the learning process, we introduce an online Dynamic Parameter Scaling (DPS) strategy. This strategy progressively integrates LoRA adapters in response to task-specific progress, enabling adaptive knowledge retention and parameter expansion. Evaluations on a diverse set of standard benchmarks (Atari, DMC, Jericho) demonstrate that ScaleZero, utilizing solely online reinforcement learning with one model, performs on par with specialized single-task agents. With the DPS strategy, it remains competitive while using just 71.5\% of the environment interactions. These findings underscore the potential of ScaleZero for effective multi-task planning. Our code is available at \textcolor{magenta}{https://github.com/opendilab/LightZero}.

Reinforcement Learning · Deep RL

Alexander Mattick, Dominik Seuß, Christopher Mutschler

Reinforcement Learning (RL) has demonstrated significant success in optimizing complex control and planning problems. However, scaling RL to real-world applications with multiple, potentially conflicting requirements requires an effective handling of constraints. We propose a novel approach to constraint satisfaction in RL algorithms, focusing on incrementally improving policy safety rather than directly projecting the policy onto a feasible region. We accomplish this by first solving a nonparametric surrogate problem which is guaranteed to contract towards the feasible set, and then cloning that solution into a neural network policy. As a result, our approach improves stability, particularly during early training stages, when the policy lacks knowledge of constraint boundaries. We provide general theoretical results guaranteeing convergence to the safe set for this class of incremental systems. Notably, even the simplest algorithm produced by our theory produces comparable or superior performance when compared to highly tuned constrained RL baselines in challenging constrained environments.

Reinforcement Learning · Deep RL

Manuel Wendl, Yarden As, Manish Prajapat, Anton Pollak, Stelian Coros, Andreas Krause

Safe exploration is a key requirement for reinforcement learning agents to learn and adapt online, beyond controlled (e.g. simulated) environments. In this work, we tackle this challenge by utilizing suboptimal yet conservative policies (e.g., obtained from offline data or simulators) as priors. Our approach, SOOPER, uses probabilistic dynamics models to optimistically explore, yet pessimistically fall back to the conservative policy prior if needed. We prove that SOOPER guarantees safety throughout learning, and establish convergence to an optimal policy by bounding its cumulative regret. Extensive experiments on key safe RL benchmarks and real-world hardware demonstrate that SOOPER is scalable, outperforms the state-of-the-art and validate our theoretical guarantees in practice.

Reinforcement Learning · Deep RL

Jongkook Heo, JAEHOON KIM, Young Jae Lee, Min Gu Kwak, Youngjoon Park, Seoung Bum Kim

Preference-based reinforcement learning (PbRL) enables agent training without explicit reward design by leveraging human feedback. Although various query sampling strategies have been proposed to improve feedback efficiency, many fail to enhance performance because they select queries from outdated experiences with low likelihood under the current policy. Such queries may no longer represent the agent's evolving behavior patterns, reducing the informativeness of human feedback. To address this issue, we propose a policy likelihood-based query sampling and critic-exploited reset (PoLiCER). Our approach uses policy likelihood-based query sampling to ensure that queries remain aligned with the agent’s evolving behavior. However, relying solely on policy-aligned sampling can result in overly localized guidance, leading to overestimation bias, as the model tends to overfit to early feedback experiences. To mitigate this, PoLiCER incorporates a dynamic resetting mechanism that selectively resets the reward estimator and its associated Q-function based on critic outputs. Experimental evaluation across diverse locomotion and robotic manipulation tasks demonstrates that PoLiCER consistently outperforms existing PbRL methods. Our code is available at https://github.com/JongKook-Heo/PoLiCER.

Deep Learning · Algorithms

Zayne Sprague, Jack Lu, Manya Wadhwa, Sedrick Keh, Mengye Ren, Greg Durrett

Reasoning models leveraging long chains of thought employ various cognitive skills, such as verification of their answers, backtracking, retrying by an alternate method, and more. Previous work has shown that when a base language model exhibits these skills, training that model further with reinforcement learning (RL) can learn to leverage them. How can we get models to leverage skills that aren't exhibited by base models? Our work, SkillFactory, is a method for fine-tuning models to roughly learn these skills during a supervised fine-tuning (SFT) stage prior to RL. Our approach does not rely on distillation from a stronger model, but instead uses samples from the model itself, rearranged to provide training data in the format of those skills. These "silver" SFT traces may be imperfect, but are nevertheless effective for priming a model to acquire skills during RL. Our evaluation shows that (1) starting from SkillFactory SFT initialization helps a model to generalize to harder variants of a task post-RL, despite lower performance pre-RL; (2) cognitive skills are indeed used by the model; (3) RLed SkillFactory models are more robust to regression on out-of-domain tasks than RLed base models. Our work suggests that inductive biases learned prior to RL help models learn robust cognitive skill use.

Reinforcement Learning · Batch/Offline

Mingxuan Cui, Duo Zhou, Yuxuan Han, Grani A. Hanasusanto, Qiong Wang, Huan Zhang, Zhengyuan Zhou

Deep reinforcement learning (RL) has achieved remarkable success, yet its deployment in real-world scenarios is often limited by vulnerability to environmental uncertainties. Distributionally robust RL (DR-RL) algorithms have been proposed to resolve this challenge, but existing approaches are largely restricted to value-based methods in tabular settings. In this work, we introduce Distributionally Robust Soft Actor-Critic (DR-SAC), the first actor–critic based DR-RL algorithm for offline learning in continuous action spaces. DR-SAC maximizes the entropy-regularized rewards against the worst possible transition models within an KL-divergence constrained uncertainty set. We derive the distributionally robust version of the soft policy iteration with a convergence guarantee and incorporate a generative modeling approach to estimate the unknown nominal transition models. Experiment results on five continuous RL tasks demonstrate our algorithm achieves up to $9.8\times$ higher average reward than the SAC baseline under common perturbations. Additionally, DR-SAC significantly improves computing efficiency and applicability to large-scale problems compared with existing DR-RL algorithms.

Deep Learning · Generative Models and Autoencoders

Philipp Becker, Niklas Freymuth, Serge Thilges, Fabian Otto, Gerhard Neumann

Reinforcement Learning (RL) with PPO-like clip objectives has become the standard choice for reward-based fine-tuning of large language models (LLMs). Although recent work has explored improved estimators of advantages and normalization, the clipping mechanism itself has remained untouched. Originally introduced as a proxy for principled KL-based trust regions, clipping is a crude approximation that often causes unstable updates and suboptimal performance. We replace the clip objective with a novel discrete differentiable trust region projection, which provides principled token-level KL constraints. The projection operates on a sparse subset of the model’s most important token logits to balance computational cost and projection effectiveness. Our approach, Trust Region Optimization for Large Language Models (TROLL), serves as a direct replacement for PPO-like clipping during training and does not alter the model’s inference behavior. Across mathematical reasoning and code generation tasks, model families, as well as advantage-estimation methods, TROLL consistently outperforms PPO-like clipping in terms of training speed, stability, and final success rates.

Deep Learning · Generative Models and Autoencoders

Feiyang Kang, Michael Kuchnik, Karthik Padthe, Marin Vlastelica, Ruoxi Jia, Carole-Jean Wu, Newsha Ardalani

In post-training for reasoning Large Language Models (LLMs), the current state of practice trains LLMs in two independent stages: Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR, shortened as "RL" below). In this work, we challenge whether high SFT scores translate to improved performance after RL. We provide extensive counter-examples where this is not true. We find high SFT scores can be biased toward simpler or more homogeneous data and are not reliably predictive of subsequent RL gains or scaled-up post-training effectiveness. In some cases, RL training on models with improved SFT performance could lead to substantially worse outcome compared to RL on the base model without SFT. We study alternative metrics and identify generalization loss on held-out reasoning examples and Pass@large k performance to provide strong proxies for the RL outcome. We trained hundreds of models up to 12B-parameter with SFT and RLVR via GRPO and ran extensive evaluations on 7 math benchmarks with up to 256 repetitions, spending $>$1M GPU hours. Experiments include models from Llama3, Mistral-Nemo, Qwen3 and multiple state-of-the-art SFT/RL datasets. Compared to directly predicting from pre-RL performance, prediction based on generalization loss and Pass@large k achieves substantial higher precision, improving $R^2$ coefficient and Spearman's rank correlation coefficient by up to 0.5 (2x). This provides strong utility for broad use cases. For example, in most experiments, we find SFT training on unique examples for a one epoch underperforms training on half examples for two epochs, either after SFT or SFT-then-RL; With the same SFT budget, training only on short examples may lead to better SFT performance, though, it often leads to worse outcome after RL compared to training on examples with varying lengths. This work develops an enhanced evaluation tool for math reasoning tasks and is open-sourced.

Theory · Reinforcement Learning and Planning

Tanmay Sadanand Ambadkar, Darshan Chudiwal, Greg Anderson, Abhinav Verma

Reinforcement learning for safety-critical tasks requires policies that are both high-performing and safe throughout the learning process. While model-predictive shielding is a promising approach, existing methods are often computationally intractable for the high-dimensional, nonlinear systems where deep RL excels, as they typically rely on a patchwork of local models. We introduce **RAMPS**, a scalable shielding framework that overcomes this limitation by leveraging a learned, linear representation of the environment's dynamics. This model can range from a linear regression in the original state space to a more complex operator learned in a high-dimensional feature space. The key is that this linear structure enables a robust, look-ahead safety technique based on a *multi-step Control Barrier Function (CBF)*. By moving beyond myopic one-step formulations, **RAMPS** accounts for model error and control delays to provide reliable, real-time interventions. The resulting framework is minimally invasive, computationally efficient, and built upon robust control-theoretic foundations. Our experiments demonstrate that **RAMPS** significantly reduces safety violations compared to existing safe RL methods while maintaining high task performance in complex control environments.

Reinforcement Learning · Deep RL

Dun Yuan, Di Wu, Xue Liu

Reinforcement learning from human feedback (RLHF) with proximal policy optimization (PPO) is widely used but often yields less diverse outputs than supervised fine-tuning, suggesting an effect in which the policy’s support contracts during on-policy optimization. We formalize this “policy contraction” with the Support Retention Ratio (SRR)—the share of SFT completions that retain non-negligible probability under the RL policy—and additionally track token-entropy, Kullback–Leibler (KL) divergence to the reference, and repetition. We propose Contraction-Aware PPO (CaPPO), a minimum-norm multi-gradient update that co-optimizes reward, entropy, and KL, paired with a controller that steers exploration toward a target token entropy. On HH-RLHF, Summarize-from-Feedback, and UltraFeedback with Qwen2-7B, Qwen2.5-14B, Mistral-7B-Instruct, and Llama-3-8B-Instruct, CaPPO increases win rate by 2 to 4 points over PPO and improves diversity, gaining 0.2 to 0.3 higher SRR. The gains persist under decoding sweeps and are robust to reward scaling and critic variance. Treating reward, diversity, and stability as first-class objectives, CaPPO mitigates contraction without sacrificing alignment performance.

Parag Dutta, Mohd Ayyoob, Shalabh Bhatnagar, Ambedkar Dukkipati

Representation learning lies at the core of deep reinforcement learning. Although CNNs have traditionally served as the primary models for encoding image observations, modifying the encoder architecture introduces challenges, especially due to the necessity of determining a new set of hyperparameters. To address this problem, we propose a powerful representation learning technique for visual reinforcement learning utilizing Fourier Neural Operators (FNO). Our findings demonstrate that the proposed FNO encoder effectively learns representations from images that encapsulate the underlying differential equations (PDEs) governing the dynamics of the environment in an online model-free RL framework. We demonstrate the applicability of our proposed architecture by replacing the CNN image encoder in PPO, A2C, and Rainbow (a Policy Gradient, Actor-Critic, and Q-Learning RL algorithm, respectively). We achieve state-of-the-art scores (in the model-free RL setting) at both the CARLA Autonomous Driving (from image observations) benchmark and the Atari 100k benchmark. Our proposed FNO encoder is compatible with all model-free reinforcement learning algorithms, enhances both rewards and sample efficiency by implicitly learning the underlying dynamics of the environment, and eliminates the need for additional hyperparameter tuning.

Suorong Yang, Peijia Li, Furao Shen, Jian Zhao

Modern deep architectures often rely on large-scale datasets, but training on these datasets incurs high computational and storage overhead. Real-world datasets often contain substantial redundancies, prompting the need for more data-efficient training paradigms. Data selection has shown promise to mitigate redundancy by identifying the most representative samples, thereby reducing training costs without compromising performance. Existing methods typically rely on static scoring metrics or pretrained models, overlooking the combined effect of selected samples and their evolving dynamics during training. To address this, we introduce the concept of epsilon-sample cover, which quantifies sample redundancy based on inter-sample relationships, capturing the intrinsic structure of the dataset. Based on this, we reformulate data selection as a reinforcement learning (RL) process, where a lightweight RL agent optimizes the selection policy by leveraging epsilon-sample cover derived from evolving dataset distribution as a reward signal. Extensive experiments across benchmark datasets and diverse architectures demonstrate that our method consistently outperforms existing state-of-the-art baselines. Models trained with our selected datasets show enhanced generalization performance with improved training efficiency.

Mingqi Yuan, Bo Li, Xin Jin, Wenjun Zeng

Hyperparameter optimization (HPO) is a billion-dollar problem in machine learning, which significantly impacts the training efficiency and model performance. However, achieving efficient and robust HPO in deep reinforcement learning (RL) is consistently challenging due to its high non-stationarity and computational cost. To tackle this problem, existing approaches attempt to adapt common HPO techniques (e.g., population-based training or Bayesian optimization) to the RL scenario. However, they remain sample-inefficient and computationally expensive, which cannot facilitate a wide range of applications. In this paper, we propose ULTHO, an ultra-lightweight yet powerful framework for fast HPO in deep RL within single runs. Specifically, we formulate the HPO process as a multi-armed bandit with clustered arms (MABC) and link it directly to long-term return optimization. ULTHO also provides a quantified and statistical perspective to filter the HPs efficiently. We test ULTHO on benchmarks including ALE, Procgen, MiniGrid, and PyBullet. Extensive experiments demonstrate that the ULTHO can achieve superior performance with a simple architecture, contributing to the development of advanced and automated RL systems.

Jinghan Yang, Zhenbo Xu, Dehua Ma, Liu Liu, Fei Liu 0008, Gong Huang, Zhaofeng He 0001

Generating accurate recipes from dish images is a challenging task that requires a deep understanding of food categories, ingredient combinations, cooking methods, and context. Current works mainly rely on the two-stage training method or supervised fine-tuning of vision-language models (VLMs). Two-stage models typically first predict ingredients from images and then generate recipes based on both ingredients and images. However, accumulated errors in ingredient prediction often lead to inaccurate recipes. Fine-tuning VLMs only fit the statistical patterns of the training data, lacking deep reasoning capabilities, which leads to severe hallucinations in the generated recipes. In this paper, we introduce a novel reinforced retrieval-augmented generation framework named RecipeRAG for recipe generation, and compare the supervised fine-tuning (SFT) paradigm and the reinforcement fine-tuning (RFT) paradigm. To effectively retrieve recipes relevant to the query image, we improve CLIP to obtain IR-CLIP as both our retriever and re-ranker by integrating metric learning and contrastive learning. The retrieved recipes are then used to enhance the generated results, improving accuracy and reducing hallucinations. However, the SFT VLM often fails to judge the quality of the retrieved recipe information and perform the complex recipe generation. Therefore, we furthermore investigate the two-phase RFT training framework. Firstly, the cold-start phase uses generated Chain-of-Thought (CoT) data for SFT to activate the reasoning capabilities of VLMs. Then, the reinforcement learning phase utilizes Group Relative Policy Optimization (GRPO) to generate multiple reasoning-answer pairs, further enhancing the generalization ability of VLMs in recipe generation tasks. Extensive evaluations on the large-scale Recipe1M dataset demonstrate that RecipeRAG outperforms all previous methods in recipe generation and exhibits strong generalization ability under the RL paradigm.

Peishuai Sun, Xiaochun Yun, Shuhao Li, Tao Yin, Chengxiang Si, Jiang Xie 0004

Deep learning-based (DL-based) malicious traffic detection models are effective but vulnerable to adversarial attacks. Existing adversarial attacks have shown promising results when targeting traffic detection models based on statistics and sequence features. However, these attacks are less effective against models that rely on payload analysis. The main reason is the difficulty in generating semantic, compliant, and functional payloads, which limits their practical application. In this paper, we propose AdvTG, an adversarial traffic generation framework to deceive DL-based malicious traffic based on the large language model (LLM) and reinforcement learning (RL). Specifically, AdvTG is designed to attack various DL-based detection models across diverse payload features and architectures, thereby enhancing the generalization capabilities of the generated adversarial traffic. Moreover, we design a specialized prompt for traffic generation tasks, where functional fields and target types are supplied as input, while non-functional fields are generated to produce the mutated traffic. This fine-tuning endows the LLM with task comprehension and traffic pattern reasoning abilities, allowing it to generate traffic that remains compliant and functional. Furthermore, leveraging RL, AdvTG automatically selects traffic fields that exhibit more robust adversarial properties. Experimental results show that AdvTG achieves over 40% attack success rate (ASR) across six detection models on four base datasets and two extended datasets, significantly outperforming other adversarial attack methods.

Ekaterina Kuzmina, Dmitrii Kriukov, Mikhail Lebedev 0001, Dmitry V. Dylov

Adaptive deep brain stimulation (aDBS) has emerged as a promising treatment for Parkinson's disease (PD). In aDBS, a surgically placed electrode sends dynamically altered stimuli to the brain based on neurophysiological feedback: an invasive gadget that limits the amount of data one could collect for optimizing the control offline. As a consequence, a plethora of synthetic models of PD and those of the control algorithms have been proposed. Herein, we introduce the first neurophysiologically realistic benchmark for comparing said models. Specifically, our methodology covers not only conventional basal ganglia circuit dynamics and pathological oscillations, but also captures 15 previously dismissed physiological attributes, such as signal instabilities and noise, neural drift, electrode conductance changes and individual variability - all modeled as spatially distributed and temporally registered features via beta-band activity in the brain and a feedback. Furthermore, we purposely built our framework as a structured environment for training and evaluating deep reinforcement learning (RL) algorithms, opening new possibilities for optimizing aDBS control strategies and inviting the machine learning community to contribute to the emerging field of intelligent neurostimulation interfaces. Code repository: https://github.com/NevVerVer/DBS-Gym

Federico Berto, Chuanbo Hua, Junyoung Park 0002, Laurin Luttmann, Yining Ma 0001, Fanchen Bu, Jiarui Wang 0002, Haoran Ye, Minsu Kim 0004, Sanghyeok Choi 等

Combinatorial optimization (CO) is fundamental to several real-world applications, from logistics and scheduling to hardware design and resource allocation. Deep reinforcement learning (RL) has recently shown significant benefits in solving CO problems, reducing reliance on domain expertise and improving computational efficiency. However, the absence of a unified benchmarking framework leads to inconsistent evaluations, limits reproducibility, and increases engineering overhead, raising barriers to adoption for new researchers. To address these challenges, we introduce RL4CO, a unified and extensive benchmark with in-depth library coverage of 27 CO problem environments and 23 state-of-the-art baselines. Built on efficient software libraries and best practices in implementation, RL4CO features modularized implementation and flexible configurations of diverse environments, policy architectures, RL algorithms, and utilities with extensive documentation. RL4CO helps researchers build on existing successes while exploring and developing their own designs, facilitating the entire research process by decoupling science from heavy engineering. We finally provide extensive benchmark studies to inspire new insights and future work. RL4CO has already attracted numerous researchers in the community and is open-sourced at https://github.com/ai4co/rl4co.

Rui Yu, Shenghua Wan, Yucen Wang, Chen-Xiao Gao, Le Gan, Zongzhang Zhang, De-Chuan Zhan

In reinforcement learning (RL), agents continually interact with the environment and use the feedback to refine their behavior. To guide policy optimization, reward models are introduced as proxies of the desired objectives, such that when the agent maximizes the accumulated reward, it also fulfills the task designer's intentions. Recently, significant attention from both academic and industrial researchers has focused on developing reward models that not only align closely with the true objectives but also facilitate policy optimization. In this survey, we provide a comprehensive review of reward modeling techniques within the RL literature. We begin by outlining the background and preliminaries in reward modeling. Next, we present an overview of recent reward modeling approaches, categorizing them based on the source, the mechanism, and the reward learning paradigm. Building on this understanding, we discuss various applications of these reward modeling techniques and review methods for evaluating reward models. Finally, we conclude by highlighting promising research directions in reward modeling. Altogether, this survey includes both established and emerging methods, filling the vacancy of a systematic review of reward models in current literature.