In standard Reinforcement Learning (RL) settings, the interaction between the agent and the environment is typically modeled as a Markov Decision Process (MDP), which assumes that the agent observes the system state instantaneously, selects an action without delay, and executes it immediately. In real-world dynamic environments, such as cyber-physical systems, this assumption often breaks down due to delays in the interaction between the agent and the system. These delays can vary stochastically over time and are typically unobservable, meaning they are unknown when deciding on an action. Existing methods deal with this uncertainty conservatively by assuming a known fixed upper bound on the delay, even if the delay is often much lower. In this work, we introduce the interaction layer, a general framework that enables agents to adaptively and seamlessly handle unobservable and time-varying delays. Specifically, the agent generates a matrix of possible future actions to handle both unpredictable delays and lost action packets sent over networks. Building on this framework, we develop a model-based algorithm, Actor-Critic with Delay Adaptation (ACDA), which dynamically adjusts to delay patterns. Our method significantly outperforms state-of-the-art approaches across a wide range of locomotion benchmark environments.
Algorithm Distillation (AD) was recently proposed as a new approach to perform In-Context Reinforcement Learning (ICRL) by modeling across-episodic training histories autoregressively with a causal transformer model. However, due to practical limitations induced by the attention mechanism, experiments were bottlenecked by the transformer's quadratic complexity and limited to simple discrete environments with short time horizons. In this work, we propose leveraging the recently proposed Selective Structured State Space Sequence (S6) models, which achieved state-of-the-art (SOTA) performance on long-range sequence modeling while scaling linearly in sequence length. Through four complex and continuous Meta Reinforcement Learning environments, we demonstrate the overall superiority of Mamba, a model built with S6 layers, over a transformer model for AD. Additionally, we show that scaling AD to very long contexts can improve ICRL performance and make it competitive even with a SOTA online meta RL baseline.
Current Reinforcement Learning (RL) methodologies for Large Language Models (LLMs) often rely on simplistic, outcome-based reward signals (e.g., final answer correctness), which limits the depth of learning from each interaction. This paper introduces Socratic Reinforcement Learning (Socratic-RL), a novel, process-oriented framework designed to address this limitation. Socratic-RL operates on the principle that deeper understanding is achieved by reflecting on the causal reasons for errors and successes within the reasoning process itself. The framework employs a decoupled "Teacher-Student" architecture, where a "Teacher AI" analyzes interaction histories, extracts causal insights, and formulates them into structured "viewpoints." These viewpoints, acting as distilled guidance, are then used by a "Student AI" to enhance its subsequent reasoning. A key innovation is the iterative self-improvement of the Teacher AI, enabling its reflective capabilities to evolve through a meta-learning loop. To manage the accumulation of knowledge, a distillation mechanism compresses learned viewpoints into the Student's parameters. By focusing on process rather than just outcome, Socratic-RL presents a pathway toward enhanced sample efficiency, superior interpretability, and a more scalable architecture for self-improving AI systems. This paper details the foundational concepts, formal mechanisms, synergies, challenges, and a concrete research roadmap for this proposed framework.
We present \textbf{FlowRL}, a novel framework for online reinforcement learning that integrates flow-based policy representation with Wasserstein-2-regularized optimization. We argue that in addition to training signals, enhancing the expressiveness of the policy class is crucial for the performance gains in RL. Flow-based generative models offer such potential, excelling at capturing complex, multimodal action distributions. However, their direct application in online RL is challenging due to a fundamental objective mismatch: standard flow training optimizes for static data imitation, while RL requires value-based policy optimization through a dynamic buffer, leading to difficult optimization landscapes. FlowRL first models policies via a state-dependent velocity field, generating actions through deterministic ODE integration from noise. We derive a constrained policy search objective that jointly maximizes Q through the flow policy while bounding the Wasserstein-2 distance to a behavior-optimal policy implicitly derived from the replay buffer. This formulation effectively aligns the flow optimization with the RL objective, enabling efficient and value-aware policy learning despite the complexity of the policy class. Empirical evaluations on DMControl and Humanoidbench demonstrate that FlowRL achieves competitive performance in online reinforcement learning benchmarks.
Deep reinforcement learning (RL) has achieved significant success, yet its application in real-world scenarios is often hindered by a lack of robustness to environmental uncertainties. To solve this challenge, some robust RL algorithms have been proposed, but most are limited to tabular settings. In this work, we propose Distributionally Robust Soft Actor-Critic (DR-SAC), a novel algorithm designed to enhance the robustness of the state-of-the-art Soft Actor-Critic (SAC) algorithm. DR-SAC aims to maximize the expected value with entropy against the worst possible transition model lying in an uncertainty set. A distributionally robust version of the soft policy iteration is derived with a convergence guarantee. For settings where nominal distributions are unknown, such as offline RL, a generative modeling approach is proposed to estimate the required nominal distributions from data. Furthermore, experimental results on a range of continuous control benchmark tasks demonstrate our algorithm achieves up to $9.8$ times the average reward of the SAC baseline under common perturbations. Additionally, compared with existing robust reinforcement learning algorithms, DR-SAC significantly improves computing efficiency and applicability to large-scale problems.
Schema linking is a critical step in Text-to-SQL task, aiming to accurately predict the table names and column names required for the SQL query based on the given question. However, current fine-tuning approaches for schema linking models employ a rote-learning paradigm, excessively optimizing for ground truth schema linking outcomes while compromising reasoning ability. This limitation arises because of the difficulty in acquiring a high-quality reasoning sample for downstream tasks. To address this, we propose Schema-R1, a reasoning schema linking model trained using reinforcement learning. Specifically, Schema-R1 consists of three key steps: constructing small batches of high-quality reasoning samples, supervised fine-tuning for cold-start initialization, and rule-based reinforcement learning training. The final results demonstrate that our method effectively enhances the reasoning ability of the schema linking model, achieving a 10\% improvement in filter accuracy compared to the existing method. Our code is available at https://github.com/hongWin/Schema-R1/.
We extend the standard reinforcement learning framework to random time horizons. While the classical setting typically assumes finite and deterministic or infinite runtimes of trajectories, we argue that multiple real-world applications naturally exhibit random (potentially trajectory-dependent) stopping times. Since those stopping times typically depend on the policy, their randomness has an effect on policy gradient formulas, which we (mostly for the first time) derive rigorously in this work both for stochastic and deterministic policies. We present two complementary perspectives, trajectory or state-space based, and establish connections to optimal control theory. Our numerical experiments demonstrate that using the proposed formulas can significantly improve optimization convergence compared to traditional approaches.
Reconfigurable Intelligent Surfaces (RISs) are composed of physical elements that can dynamically alter electromagnetic wave properties to enhance beamforming and leading to improvements in areas with low coverage properties. They have the potential to be combined with Reinforcement Learning (RL) techniques to achieve network performance and energy efficiency via optimization techniques. In addition to performance and energy improvements, it is also crucial to consider the concept of fair communications. RISs must ensure that User Equipment (UE) units receive their signals with adequate strength, without other UE being deprived of service due to insufficient power. In this paper, we address such a problem. We explore the fairness properties of previous work and propose a novel method that aims at obtaining an efficient and fair duplex RIS-RL system for multiple legitimate UE units. We report and discuss our experimental work and simulation results. We also release our code and datasets to foster further research in the topic.
Discrete choice modelling is a theory-driven modelling framework for understanding and forecasting choice behaviour. To obtain behavioural insights, modellers test several competing model specifications in their attempts to discover the 'true' data generation process. This trial-and-error process requires expertise, is time-consuming, and relies on subjective theoretical assumptions. Although metaheuristics have been proposed to assist choice modellers, they treat model specification as a classic optimisation problem, relying on static strategies, applying predefined rules, and neglecting outcomes from previous estimated models. As a result, current metaheuristics struggle to prioritise promising search regions, adapt exploration dynamically, and transfer knowledge to other modelling tasks. To address these limitations, we introduce a deep reinforcement learning-based framework where an 'agent' specifies models by estimating them and receiving rewards based on goodness-of-fit and parsimony. Results demonstrate the agent dynamically adapts its strategies to identify promising specifications across data generation processes, showing robustness and potential transferability, without prior domain knowledge.
Purpose: This study aims to address the challenges of controlling unstable and nonlinear systems by proposing an adaptive PID controller based on predictive reinforcement learning (PRL-PID), where the PRL-PID combines the advantages of both data-driven and model-driven approaches. Design/methodology/approach: A predictive reinforcement learning framework is introduced, incorporating action smooth strategy to suppress overshoot and oscillations, and a hierarchical reward function to support training. Findings: Experimental results show that the PRL-PID controller achieves superior stability and tracking accuracy in nonlinear, unstable, and strongly coupled systems, consistently outperforming existing RL-tuned PID methods while maintaining excellent robustness and adaptability across diverse operating conditions. Originality/Value: By adopting predictive learning, the proposed PRL-PID integrates system model priors into data-driven control, enhancing both the control framework's training efficiency and the controller's stability. As a result, PRL-PID provides a balanced blend of model-based and data-driven approaches, delivering robust, high-performance control.
Recent advances in Vision-Language-Action (VLA) model have demonstrated strong generalization capabilities across diverse scenes, tasks, and robotic platforms when pretrained at large-scale datasets. However, these models still require task-specific fine-tuning in novel environments, a process that relies almost exclusively on supervised fine-tuning (SFT) using static trajectory datasets. Such approaches neither allow robot to interact with environment nor do they leverage feedback from live execution. Also, their success is critically dependent on the size and quality of the collected trajectories. Reinforcement learning (RL) offers a promising alternative by enabling closed-loop interaction and aligning learned policies directly with task objectives. In this work, we draw inspiration from the ideas of GRPO and propose the Trajectory-wise Group Relative Policy Optimization (TGRPO) method. By fusing step-level and trajectory-level advantage signals, this method improves GRPO's group-level advantage estimation, thereby making the algorithm more suitable for online reinforcement learning training of VLA. Experimental results on ten manipulation tasks from the libero-object benchmark demonstrate that TGRPO consistently outperforms various baseline methods, capable of generating more robust and efficient policies across multiple tested scenarios. Our source codes are available at: https://github.com/hahans/TGRPO
Reinforcement learning with verifiable rewards (RLVR) has become a key technique for enhancing large language models (LLMs), with verification engineering playing a central role. However, best practices for RL in instruction following remain underexplored. In this work, we explore the verification challenge in RL for instruction following and propose VerIF, a verification method that combines rule-based code verification with LLM-based verification from a large reasoning model (e.g., QwQ-32B). To support this approach, we construct a high-quality instruction-following dataset, VerInstruct, containing approximately 22,000 instances with associated verification signals. We apply RL training with VerIF to two models, achieving significant improvements across several representative instruction-following benchmarks. The trained models reach state-of-the-art performance among models of comparable size and generalize well to unseen constraints. We further observe that their general capabilities remain unaffected, suggesting that RL with VerIF can be integrated into existing RL recipes to enhance overall model performance. We have released our datasets, codes, and models to facilitate future research at https://github.com/THU-KEG/VerIF.
Sample efficiency and exploration remain critical challenges in Deep Reinforcement Learning (DRL), particularly in complex domains. Offline RL, which enables agents to learn optimal policies from static, pre-collected datasets, has emerged as a promising alternative. However, offline RL is constrained by issues such as out-of-distribution (OOD) actions that limit policy performance and generalization. To overcome these limitations, we propose Meta Offline-Online Reinforcement Learning (MOORL), a hybrid framework that unifies offline and online RL for efficient and scalable learning. While previous hybrid methods rely on extensive design components and added computational complexity to utilize offline data effectively, MOORL introduces a meta-policy that seamlessly adapts across offline and online trajectories. This enables the agent to leverage offline data for robust initialization while utilizing online interactions to drive efficient exploration. Our theoretical analysis demonstrates that the hybrid approach enhances exploration by effectively combining the complementary strengths of offline and online data. Furthermore, we demonstrate that MOORL learns a stable Q-function without added complexity. Extensive experiments on 28 tasks from the D4RL and V-D4RL benchmarks validate its effectiveness, showing consistent improvements over state-of-the-art offline and hybrid RL baselines. With minimal computational overhead, MOORL achieves strong performance, underscoring its potential for practical applications in real-world scenarios.
We study reinforcement learning from human feedback in general Markov decision processes, where agents learn from trajectory-level preference comparisons. A central challenge in this setting is to design algorithms that select informative preference queries to identify the underlying reward while ensuring theoretical guarantees. We propose a meta-algorithm based on randomized exploration, which avoids the computational challenges associated with optimistic approaches and remains tractable. We establish both regret and last-iterate guarantees under mild reinforcement learning oracle assumptions. To improve query complexity, we introduce and analyze an improved algorithm that collects batches of trajectory pairs and applies optimal experimental design to select informative comparison queries. The batch structure also enables parallelization of preference queries, which is relevant in practical deployment as feedback can be gathered concurrently. Empirical evaluation confirms that the proposed method is competitive with reward-based reinforcement learning while requiring a small number of preference queries.
Combinatorial optimization problems are notoriously challenging due to their discrete structure and exponentially large solution space. Recent advances in deep reinforcement learning (DRL) have enabled the learning heuristics directly from data. However, DRL methods often suffer from limited exploration and susceptibility to local optima. On the other hand, evolutionary algorithms such as Genetic Algorithms (GAs) exhibit strong global exploration capabilities but are typically sample inefficient and computationally intensive. In this work, we propose the Evolutionary Augmentation Mechanism (EAM), a general and plug-and-play framework that synergizes the learning efficiency of DRL with the global search power of GAs. EAM operates by generating solutions from a learned policy and refining them through domain-specific genetic operations such as crossover and mutation. These evolved solutions are then selectively reinjected into the policy training loop, thereby enhancing exploration and accelerating convergence. We further provide a theoretical analysis that establishes an upper bound on the KL divergence between the evolved solution distribution and the policy distribution, ensuring stable and effective policy updates. EAM is model-agnostic and can be seamlessly integrated with state-of-the-art DRL solvers such as the Attention Model, POMO, and SymNCO. Extensive results on benchmark problems (e.g., TSP, CVRP, PCTSP, and OP) demonstrate that EAM significantly improves both solution quality and training efficiency over competitive baselines.
We introduce a novel task of clustering trajectories from offline reinforcement learning (RL) datasets, where each cluster center represents the policy that generated its trajectories. By leveraging the connection between the KL-divergence of offline trajectory distributions and a mixture of policy-induced distributions, we formulate a natural clustering objective. To solve this, we propose Policy-Guided K-means (PG-Kmeans) and Centroid-Attracted Autoencoder (CAAE). PG-Kmeans iteratively trains behavior cloning (BC) policies and assigns trajectories based on policy generation probabilities, while CAAE resembles the VQ-VAE framework by guiding the latent representations of trajectories toward the vicinity of specific codebook entries to achieve clustering. Theoretically, we prove the finite-step convergence of PG-Kmeans and identify a key challenge in offline trajectory clustering: the inherent ambiguity of optimal solutions due to policy-induced conflicts, which can result in multiple equally valid but structurally distinct clusterings. Experimentally, we validate our methods on the widely used D4RL dataset and custom GridWorld environments. Our results show that both PG-Kmeans and CAAE effectively partition trajectories into meaningful clusters. They offer a promising framework for policy-based trajectory clustering, with broad applications in offline RL and beyond.
Despite a strong desire to quit, individuals with long-term substance use disorder (SUD) often struggle to resist drug use, even when aware of its harmful consequences. This disconnect between knowledge and compulsive behavior reflects a fundamental cognitive-behavioral conflict in addiction. Neurobiologically, differential cue-induced activity within striatal subregions, along with dopamine-mediated connectivity from the ventral to the dorsal striatum, contributes to compulsive drug-seeking. However, the functional mechanism linking these findings to behavioral conflict remains unclear. Another hallmark of addiction is temporal discounting: individuals with drug dependence exhibit steeper discount rates than non-users. Assuming the ventral-dorsal striatal organization reflects a gradient from cognitive to motor representations, addiction can be modeled within a hierarchical reinforcement learning (HRL) framework. However, integrating discounting into biologically grounded HRL remains an open challenge. In this work, we build on a model showing how action choices reinforced with drug rewards become insensitive to the negative consequences that follow. We address the integration of discounting by ensuring natural reward values converge across all levels in the HRL hierarchy, while drug rewards diverge due to their dopaminergic effects. Our results show that high discounting amplifies drug-seeking across the hierarchy, linking faster discounting with increased addiction severity and impulsivity. We demonstrate alignment with empirical findings on temporal discounting and propose testable predictions, establishing addiction as a disorder of hierarchical decision-making.
Reinforcement learning (RL) algorithms can find an optimal policy for a single agent to accomplish a particular task. However, many real-world problems require multiple agents to collaborate in order to achieve a common goal. For example, a robot executing a task in a warehouse may require the assistance of a drone to retrieve items from high shelves. In Decentralized Multi-Agent RL (DMARL), agents learn independently and then combine their policies at execution time, but often must satisfy constraints on compatibility of local policies to ensure that they can achieve the global task when combined. In this paper, we study how providing high-level symbolic knowledge to agents can help address unique challenges of this setting, such as privacy constraints, communication limitations, and performance concerns. In particular, we extend the formal tools used to check the compatibility of local policies with the team task, making decentralized training with theoretical guarantees usable in more scenarios. Furthermore, we empirically demonstrate that symbolic knowledge about the temporal evolution of events in the environment can significantly expedite the learning process in DMARL.
We study the problem of sample efficient reinforcement learning, where prior data such as demonstrations are provided for initialization in lieu of a dense reward signal. A natural approach is to incorporate an imitation learning objective, either as regularization during training or to acquire a reference policy. However, imitation learning objectives can ultimately degrade long-term performance, as it does not directly align with reward maximization. In this work, we propose to use prior data solely for guiding exploration via noise added to the policy, sidestepping the need for explicit behavior cloning constraints. The key insight in our framework, Data-Guided Noise (DGN), is that demonstrations are most useful for identifying which actions should be explored, rather than forcing the policy to take certain actions. Our approach achieves up to 2-3x improvement over prior reinforcement learning from offline data methods across seven simulated continuous control tasks.
Protein generative models have shown remarkable promise in protein design but still face limitations in success rate, due to the scarcity of high-quality protein datasets for supervised pretraining. We present ProteinZero, a novel framework that enables scalable, automated, and continuous self-improvement of the inverse folding model through online reinforcement learning. To achieve computationally tractable online feedback, we introduce efficient proxy reward models based on ESM-fold and a novel rapid ddG predictor that significantly accelerates evaluation speed. ProteinZero employs a general RL framework balancing multi-reward maximization, KL-divergence from a reference model, and a novel protein-embedding level diversity regularization that prevents mode collapse while promoting higher sequence diversity. Through extensive experiments, we demonstrate that ProteinZero substantially outperforms existing methods across every key metric in protein design, achieving significant improvements in structural accuracy, designability, thermodynamic stability, and sequence diversity. Most impressively, ProteinZero reduces design failure rates by approximately 36% - 48% compared to widely-used methods like ProteinMPNN, ESM-IF and InstructPLM, consistently achieving success rates exceeding 90% across diverse and complex protein folds. Notably, the entire RL run on CATH-4.3 can be done with a single 8 X GPU node in under 3 days, including reward computation. Our work establishes a new paradigm for protein design where models evolve continuously from their own generated outputs, opening new possibilities for exploring the vast protein design space.