Understanding how to efficiently learn while adhering to safety constraints is essential for using online reinforcement learning in practical applications. However, proving rigorous regret bounds for safety-constrained reinforcement learning is difficult due to the complex interaction between safety, exploration, and exploitation. In this work, we seek to establish foundations for safety-constrained reinforcement learning by studying the canonical problem of controlling a one-dimensional linear dynamical system with unknown dynamics. We study the safety-constrained version of this problem, where the state must with high probability stay within a safe region, and we provide the first safe algorithm that achieves regret of $\tilde{O}_T(\sqrt{T})$. Furthermore, the regret is with respect to the baseline of truncated linear controllers, a natural baseline of non-linear controllers that are well-suited for safety-constrained linear systems. In addition to introducing this new baseline, we also prove several desirable continuity properties of the optimal controller in this baseline. In showing our main result, we prove that whenever the constraints impact the optimal controller, the non-linearity of our controller class leads to a faster rate of learning than in the unconstrained setting.
Recently, foundation models such as OpenAI's O1 and O3, along with DeepSeek's R1, have demonstrated strong reasoning capacities and problem-solving skills acquired through large-scale reinforcement learning (RL), with wide applications in mathematics, coding, science, intelligent agents, and virtual assistants. In this work, we introduce an off-policy reinforcement learning algorithm, EM Policy Gradient, aimed at enhancing LLM reasoning by optimizing expected return over reasoning trajectories. We frame the reasoning task as an Expectation-Maximization (EM) optimization problem, alternating between sampling diverse rationale trajectories and performing reward-guided fine-tuning. Unlike PPO and GRPO, which rely on complex importance weights and heuristic clipping, our method provides a simpler, more principled off-policy policy gradient approach, eliminating these complexities while maintaining strong performance. We evaluate the effectiveness of EM Policy Gradient on the GSM8K and MATH (HARD) datasets, where it achieves performance comparable to or slightly surpassing the state-of-the-art GRPO, while offering additional advantages in scalability, simplicity, and reasoning conciseness. Moreover, models fine-tuned with our method exhibit cognitive behaviors, such as sub-problem decomposition, self-verification, and backtracking, highlighting its potential to enhance both the interpretability and robustness of LLM reasoning.
This paper addresses the challenge of offline policy learning in reinforcement learning with continuous action spaces when unmeasured confounders are present. While most existing research focuses on policy evaluation within partially observable Markov decision processes (POMDPs) and assumes discrete action spaces, we advance this field by establishing a novel identification result to enable the nonparametric estimation of policy value for a given target policy under an infinite-horizon framework. Leveraging this identification, we develop a minimax estimator and introduce a policy-gradient-based algorithm to identify the in-class optimal policy that maximizes the estimated policy value. Furthermore, we provide theoretical results regarding the consistency, finite-sample error bound, and regret bound of the resulting optimal policy. Extensive simulations and a real-world application using the German Family Panel data demonstrate the effectiveness of our proposed methodology.
Recent work shows that reinforcement learning(RL) can markedly sharpen the reasoning ability of large language models (LLMs) by prompting them to "think before answering." Yet whether and how these gains transfer to audio-language reasoning remains largely unexplored. We extend the Group-Relative Policy Optimization (GRPO) framework from DeepSeek-R1 to a Large Audio-Language Model (LALM), and construct a 32k sample multiple-choice corpus. Using a two-stage regimen supervised fine-tuning on structured and unstructured chains-of-thought, followed by curriculum-guided GRPO, we systematically compare implicit vs. explicit, and structured vs. free form reasoning under identical architectures. Our structured audio reasoning model, SARI (Structured Audio Reasoning via Curriculum-Guided Reinforcement Learning), achieves a 16.35% improvement in average accuracy over the base model Qwen2-Audio-7B-Instruct. Furthermore, the variant built upon Qwen2.5-Omni reaches state-of-the-art performance of 67.08% on the MMAU test-mini benchmark. Ablation experiments show that on the base model we use: (i) SFT warm-up is important for stable RL training, (ii) structured chains yield more robust generalization than unstructured ones, and (iii) easy-to-hard curricula accelerate convergence and improve final performance. These findings demonstrate that explicit, structured reasoning and curriculum learning substantially enhances audio-language understanding.
This work presents an overview of the technical details behind a high performance reinforcement learning policy deployment with the Spot RL Researcher Development Kit for low level motor access on Boston Dynamics Spot. This represents the first public demonstration of an end to end end reinforcement learning policy deployed on Spot hardware with training code publicly available through Nvidia IsaacLab and deployment code available through Boston Dynamics. We utilize Wasserstein Distance and Maximum Mean Discrepancy to quantify the distributional dissimilarity of data collected on hardware and in simulation to measure our sim2real gap. We use these measures as a scoring function for the Covariance Matrix Adaptation Evolution Strategy to optimize simulated parameters that are unknown or difficult to measure from Spot. Our procedure for modeling and training produces high quality reinforcement learning policies capable of multiple gaits, including a flight phase. We deploy policies capable of over 5.2ms locomotion, more than triple Spots default controller maximum speed, robustness to slippery surfaces, disturbance rejection, and overall agility previously unseen on Spot. We detail our method and release our code to support future work on Spot with the low level API.
We introduce Seed1.5-Thinking, capable of reasoning through thinking before responding, resulting in improved performance on a wide range of benchmarks. Seed1.5-Thinking achieves 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA, demonstrating excellent reasoning abilities in STEM and coding. Beyond reasoning tasks, the method demonstrates notable generalization across diverse domains. For instance, it surpasses DeepSeek R1 by 8% in win rate on non-reasoning tasks, indicating its broader applicability. Compared to other state-of-the-art reasoning models, Seed1.5-Thinking is a Mixture-of-Experts (MoE) model with a relatively small size, featuring 20B activated and 200B total parameters. As part of our effort to assess generalized reasoning, we develop two internal benchmarks, BeyondAIME and Codeforces, both of which will be publicly released to support future research. Model trial link: https://www.volcengine.com/experience/ark.
Reinforcement learning (RL) algorithms, due to their reliance on external systems to learn from, require digital environments (e.g., simulators) with very simple interfaces, which in turn constrain significantly the implementation of such environments. In particular, these environments are implemented either as separate processes or as state machines, leading to synchronization and communication overheads in the first case, and to unstructured programming in the second. We propose a new domain-specific, co-routine-based, compiled language, called Rulebook, designed to automatically generate the state machine required to interact with machine learning (ML) algorithms and similar applications, with no performance overhead. Rulebook allows users to express programs without needing to be aware of the specific interface required by the ML components. By decoupling the execution model of the program from the syntactical encoding of the program, and thus without the need for manual state management, Rulebook allows to create larger and more sophisticated environments at a lower development cost.
This work reframes the Text-to-SQL task as a pathway for teaching large language models (LLMs) to reason over and manipulate tabular data--moving beyond the traditional focus on query generation. We propose a two-stage framework that leverages SQL supervision to develop transferable table reasoning capabilities. First, we synthesize detailed chain-of-thought (CoT) traces from real-world SQL queries, providing step-by-step, clause-level supervision that teaches the model how to traverse, filter, and aggregate table fields. Second, we introduce a Group Relative Policy Optimization (GRPO) reinforcement learning objective that connects SQL execution accuracy to generalizable reasoning by encouraging steps that extend beyond task-specific syntax and transfer across datasets. Empirically, our approach improves performance on standard Text-to-SQL benchmarks and achieves substantial gains on reasoning-intensive datasets such as BIRD and CRT-QA, demonstrating enhanced generalization and interpretability. Specifically, the distilled-quantized LLaMA model achieved a relative 33.9\% increase in accuracy when trained on Text-to-SQL tasks, while Qwen achieved a relative 14.5\% increase. These results suggest that SQL can serve not only as a target formalism but also as an effective scaffold for learning robust, transferable reasoning over structured data.
Recently, deep reinforcement learning has emerged as a promising approach for solving complex combinatorial optimization problems. Broadly, deep reinforcement learning methods fall into two categories: policy-based and value-based. While value-based approaches have achieved notable success in domains such as the Arcade Learning Environment, the combinatorial optimization community has predominantly favored policy-based methods, often overlooking the potential of value-based algorithms. In this work, we conduct a comprehensive empirical evaluation of value-based algorithms, including the deep q-network and several of its advanced extensions, within the context of two complex combinatorial problems: the job-shop and the flexible job-shop scheduling problems, two fundamental challenges with multiple industrial applications. Our results challenge the assumption that policy-based methods are inherently superior for combinatorial optimization. We show that several value-based approaches can match or even outperform the widely adopted proximal policy optimization algorithm, suggesting that value-based strategies deserve greater attention from the combinatorial optimization community. Our code is openly available at: https://github.com/AJ-Correa/Unraveling-the-Rainbow.
Reinforcement learning from expert demonstrations has long remained a challenging research problem, and existing state-of-the-art methods using behavioral cloning plus further RL training often suffer from poor generalization, low sample efficiency, and poor model interpretability. Inspired by the strong reasoning abilities of large language models (LLMs), we propose a novel strategy-based reinforcement learning framework integrated with LLMs called DYnamic STrategy Induction with Llms for reinforcement learning (DYSTIL) to overcome these limitations. DYSTIL dynamically queries a strategy-generating LLM to induce textual strategies based on advantage estimations and expert demonstrations, and gradually internalizes induced strategies into the RL agent through policy optimization to improve its performance through boosting policy generalization and enhancing sample efficiency. It also provides a direct textual channel to observe and interpret the evolution of the policy's underlying strategies during training. We test DYSTIL over challenging RL environments from Minigrid and BabyAI, and empirically demonstrate that DYSTIL significantly outperforms state-of-the-art baseline methods by 17.75% in average success rate while also enjoying higher sample efficiency during the learning process.
We present a reinforcement learning strategy for use in multi-agent foraging systems in which the learning is centralised to a single agent and its model is periodically disseminated among the population of non-learning agents. In a domain where multi-agent reinforcement learning (MARL) is the common approach, this approach aims to significantly reduce the computational and energy demands compared to approaches such as MARL and centralised learning models. By developing high performing foraging agents, these approaches can be translated into real-world applications such as logistics, environmental monitoring, and autonomous exploration. A reward function was incorporated into this approach that promotes role development among agents, without explicit directives. This led to the differentiation of behaviours among the agents. The implicit encouragement of role differentiation allows for dynamic actions in which agents can alter roles dependent on their interactions with the environment without the need for explicit communication between agents.
Multi-Agent Reinforcement Learning (MARL) has become a powerful framework for numerous real-world applications, modeling distributed decision-making and learning from interactions with complex environments. Resource Allocation Optimization (RAO) benefits significantly from MARL's ability to tackle dynamic and decentralized contexts. MARL-based approaches are increasingly applied to RAO challenges across sectors playing pivotal roles to Industry 4.0 developments. This survey provides a comprehensive review of recent MARL algorithms for RAO, encompassing core concepts, classifications, and a structured taxonomy. By outlining the current research landscape and identifying primary challenges and future directions, this survey aims to support researchers and practitioners in leveraging MARL's potential to advance resource allocation solutions.
In this paper, a novel quantum classical hybrid framework is proposed that synergizes quantum with Classical Reinforcement Learning. By leveraging the inherent parallelism of quantum computing, the proposed approach generates robust Q tables and specialized turn cost estimations, which are then integrated with a classical Reinforcement Learning pipeline. The Classical Quantum fusion results in rapid convergence of training, reducing the training time significantly and improved adaptability in scenarios featuring static, dynamic, and moving obstacles. Simulator based evaluations demonstrate significant enhancements in path efficiency, trajectory smoothness, and mission success rates, underscoring the potential of framework for real time, autonomous navigation in complex and unpredictable environments. Furthermore, the proposed framework was tested beyond simulations on practical scenarios, including real world map data such as the IIT Delhi campus, reinforcing its potential for real time, autonomous navigation in complex and unpredictable environments.
To address the challenges of high resource dynamism and intensive task concurrency in microservice systems, this paper proposes an adaptive resource scheduling method based on the A3C reinforcement learning algorithm. The scheduling problem is modeled as a Markov Decision Process, where policy and value networks are jointly optimized to enable fine-grained resource allocation under varying load conditions. The method incorporates an asynchronous multi-threaded learning mechanism, allowing multiple agents to perform parallel sampling and synchronize updates to the global network parameters. This design improves both policy convergence efficiency and model stability. In the experimental section, a real-world dataset is used to construct a scheduling scenario. The proposed method is compared with several typical approaches across multiple evaluation metrics, including task delay, scheduling success rate, resource utilization, and convergence speed. The results show that the proposed method delivers high scheduling performance and system stability in multi-task concurrent environments. It effectively alleviates the resource allocation bottlenecks faced by traditional methods under heavy load, demonstrating its practical value for intelligent scheduling in microservice systems.
Vehicular Ad-hoc Networks (VANETs) are integral to intelligent transportation systems, enabling vehicles to offload computational tasks to nearby roadside units (RSUs) and mobile edge computing (MEC) servers for real-time processing. However, the highly dynamic nature of VANETs introduces challenges, such as unpredictable network conditions, high latency, energy inefficiency, and task failure. This research addresses these issues by proposing a hybrid AI framework that integrates supervised learning, reinforcement learning, and Particle Swarm Optimization (PSO) for intelligent task offloading and resource allocation. The framework leverages supervised models for predicting optimal offloading strategies, reinforcement learning for adaptive decision-making, and PSO for optimizing latency and energy consumption. Extensive simulations demonstrate that the proposed framework achieves significant reductions in latency and energy usage while improving task success rates and network throughput. By offering an efficient, and scalable solution, this framework sets the foundation for enhancing real-time applications in dynamic vehicular environments.
Safe Reinforcement Learning (Safe RL) aims to ensure safety when an RL agent conducts learning by interacting with real-world environments where improper actions can induce high costs or lead to severe consequences. In this paper, we propose a novel Safe Skill Planning (SSkP) approach to enhance effective safe RL by exploiting auxiliary offline demonstration data. SSkP involves a two-stage process. First, we employ PU learning to learn a skill risk predictor from the offline demonstration data. Then, based on the learned skill risk predictor, we develop a novel risk planning process to enhance online safe RL and learn a risk-averse safe policy efficiently through interactions with the online RL environment, while simultaneously adapting the skill risk predictor to the environment. We conduct experiments in several benchmark robotic simulation environments. The experimental results demonstrate that the proposed approach consistently outperforms previous state-of-the-art safe RL methods.
Modular self-reconfigurable satellites refer to satellite clusters composed of individual modular units capable of altering their configurations. The configuration changes enable the execution of diverse tasks and mission objectives. Existing path planning algorithms for reconfiguration often suffer from high computational complexity, poor generalization capability, and limited support for diverse target configurations. To address these challenges, this paper proposes a goal-oriented reinforcement learning-based path planning algorithm. This algorithm is the first to address the challenge that previous reinforcement learning methods failed to overcome, namely handling multiple target configurations. Moreover, techniques such as Hindsight Experience Replay and Invalid Action Masking are incorporated to overcome the significant obstacles posed by sparse rewards and invalid actions. Based on these designs, our model achieves a 95% and 73% success rate in reaching arbitrary target configurations in a modular satellite cluster composed of four and six units, respectively.
This paper presents Tunnel, a simple, open source, reinforcement learning training environment for high performance aircraft. It integrates the F16 3D nonlinear flight dynamics into OpenAI Gymnasium python package. The template includes primitives for boundaries, targets, adversaries and sensing capabilities that may vary depending on operational need. This offers mission planners a means to rapidly respond to evolving environments, sensor capabilities and adversaries for autonomous air combat aircraft. It offers researchers access to operationally relevant aircraft physics. Tunnel code base is accessible to anyone familiar with Gymnasium and/or those with basic python skills. This paper includes a demonstration of a week long trade study that investigated a variety of training methods, observation spaces, and threat presentations. This enables increased collaboration between researchers and mission planners which can translate to a national military advantage. As warfare becomes increasingly reliant upon automation, software agility will correlate with decision advantages. Airmen must have tools to adapt to adversaries in this context. It may take months for researchers to develop skills to customize observation, actions, tasks and training methodologies in air combat simulators. In Tunnel, this can be done in a matter of days.
Conditional decision generation with diffusion models has shown powerful competitiveness in reinforcement learning (RL). Recent studies reveal the relation between energy-function-guidance diffusion models and constrained RL problems. The main challenge lies in estimating the intermediate energy, which is intractable due to the log-expectation formulation during the generation process. To address this issue, we propose the Analytic Energy-guided Policy Optimization (AEPO). Specifically, we first provide a theoretical analysis and the closed-form solution of the intermediate guidance when the diffusion model obeys the conditional Gaussian transformation. Then, we analyze the posterior Gaussian distribution in the log-expectation formulation and obtain the target estimation of the log-expectation under mild assumptions. Finally, we train an intermediate energy neural network to approach the target estimation of log-expectation formulation. We apply our method in 30+ offline RL tasks to demonstrate the effectiveness of our method. Extensive experiments illustrate that our method surpasses numerous representative baselines in D4RL offline reinforcement learning benchmarks.
The performance of Offline reinforcement learning is significantly impacted by the issue of state distributional shift, and out-of-distribution (OOD) state correction is a popular approach to address this problem. In this paper, we propose a novel method named Density-Aware Safety Perception (DASP) for OOD state correction. Specifically, our method encourages the agent to prioritize actions that lead to outcomes with higher data density, thereby promoting its operation within or the return to in-distribution (safe) regions. To achieve this, we optimize the objective within a variational framework that concurrently considers both the potential outcomes of decision-making and their density, thus providing crucial contextual information for safe decision-making. Finally, we validate the effectiveness and feasibility of our proposed method through extensive experimental evaluations on the offline MuJoCo and AntMaze suites.