Reinforcement Learning

Reinforcement Learning with Verifiable Rewards (RLVR) is widely used for post-training Large Language Models, but practical verifiers can make errors. We study how the rate and structure of reward noise affect RLVR in code generation, with a preliminary scientific-reasoning check. In multi-seed Qwen3 8B experiments on MBPP, mean validation reward over two fixed late evaluations is within 1 percentage point of the clean baseline at the tested resampled group-rollout noise rates through 20%, and within about 2 points at 30%. Confidence intervals allow larger losses; these point estimates do not establish a general tolerance threshold. We also examine full-program pass@k, four controlled noise structures, two model-based verifiers, and policy models from three families spanning 4B-9B parameters. We derive conditional advantage distributions for symmetric and asymmetric group noise, including retained format penalties, and show why clipping limits simple gradient-scaling arguments. The analysis identifies information preserved by whole-group corruption and limits on interpreting our asymmetric sweep as a precision-recall comparison. Overall, the results indicate that imperfect verification can support effective RLVR in the tested settings, while aggregate error rates alone do not characterize the learning signal.

Time Slotted Channel Hopping (TSCH) is a widely adopted Media Access Control (MAC) protocol within the IEEE 802.15.4e standard, designed to provide reliable and energy-efficient communication in Industrial Internet of Things (IIoT) networks. However, state-of-the-art TSCH schedulers rely on static slot allocations, resulting in idle listening and unnecessary power consumption under dynamic traffic conditions. This paper introduces RL-ASL, a reinforcement learning-driven adaptive listening framework that dynamically decides whether to activate or skip a scheduled listening slot based on real-time network conditions. By integrating learning-based slot skipping with standard TSCH scheduling, RL-ASL reduces idle listening while preserving synchronization and delivery reliability. Experimental results on the FIT IoT-LAB testbed and Cooja network simulator show that RL-ASL achieves up to 46% lower power consumption than baseline scheduling protocols, while maintaining near-perfect reliability and reducing average latency by up to 96% compared to PRIL-M. Its link-based variant, RL-ASL-LB, further improves delay performance under high contention with similar energy efficiency. Importantly, RL-ASL performs inference on constrained motes with negligible overhead, as model training is fully performed offline. Overall, RL-ASL provides a practical, scalable, and energy-aware scheduling mechanism for next-generation low-power IIoT networks.

Safety in reinforcement learning (RL) is typically enforced through objective shaping while keeping environment dynamics stationary with respect to observable state-action pairs. Under delayed harm, this can lead to replay: after a washout period, reintroducing the same stimulus under matched observable conditions reproduces a similar harmful cascade. We introduce the Replay Suppression Diagnostic (RSD), a controlled exposure-decay-replay protocol that isolates this failure mode under frozen-policy evaluation. We show that, under stationary observable transition kernels, replay cannot be structurally suppressed without inducing a persistent shift in replay-time action distributions. Motivated by platform-mediated systems, we propose Regret-Aware Policy Optimization (RAPO), which augments the environment with persistent harm-trace and scar fields and applies a bounded, mass-preserving transition reweighting to reduce reachability of historically harmful regions. On graph diffusion tasks (50-1000 nodes), RAPO suppresses replay, reducing re-amplification gain (RAG) from 0.98 to 0.33 on 250-node graphs while retaining 82% of task return. Disabling transition deformation only during replay restores re-amplification (RAG 0.91), isolating environment-level deformation as the causal mechanism.

Deep reinforcement learning has recently achieved strong results in quadrupedal locomotion, yet policies trained in simulation often fail to transfer when the environment changes. Evolutionary reinforcement learning aims to address this limitation by combining gradient-based policy optimisation with population-driven exploration. This work evaluates four methods on a simulated walking task: DDPG, TD3, and two Cross-Entropy-based variants CEM-DDPG and CEM-TD3. All agents are trained on flat terrain and later tested both on this domain and on a rough terrain not encountered during training. TD3 performs best among the standard deep RL baselines on flat ground with a mean reward of 5927.26, while CEM-TD3 achieves the highest rewards overall during training and evaluation 17611.41. Under the rough-terrain transfer test, performance of the deep RL methods drops sharply. DDPG achieves -1016.32 and TD3 achieves -99.73, whereas the evolutionary variants retain much of their capability. CEM-TD3 records the strongest transfer performance with a mean reward of 19574.33. These findings suggest that incorporating evolutionary search can reduce overfitting and improve policy robustness in locomotion tasks, particularly when deployment conditions differ from those seen during training.

A key challenge in scaling up Reinforcement Learning is generalizing learned behaviour. Without the ability to carry forward acquired knowledge an agent is doomed to learn each task from scratch. In this paper we develop a new formalism for transfer by virtue of state abstraction. Based on task-independent, compact observations (outcomes) of the environment, we introduce Outcome-Predictive State Representations (OPSRs), agent-centered and task-independent abstractions that are made up of predictions of outcomes. We show formally and empirically that they have the potential for optimal but limited transfer, then overcome this trade-off by introducing OPSR-based skills, i.e. abstract actions (based on options) that can be reused between tasks as a result of state abstraction. In a series of empirical studies, we learn OPSR-based skills from demonstrations and show how they speed up learning considerably in entirely new and unseen tasks without any pre-processing. We believe that the framework introduced in this work is a promising step towards transfer in RL in general, and towards transfer through combining state and action abstraction specifically.

Energy efficiency in mobile networks is crucial for sustainable telecommunications infrastructure, particularly as network densification continues to increase power consumption. Sleep mechanisms for the components in mobile networks can reduce energy use, but deciding which components to put to sleep, when, and for how long while preserving quality of service (QoS) remains a difficult optimisation problem. In this paper, we utilise reinforcement learning with reward machines (RMs) to make sleep-control decisions that balance immediate energy savings and long-term QoS impact, i.e. time-averaged packet drop rates for deadline-constrained traffic and time-averaged minimum-throughput guarantees for constant-rate users. A challenge is that time-averaged constraints depend on cumulative performance over time rather than immediate performance. As a result, the effective reward is non-Markovian, and optimal actions depend on operational history rather than the instantaneous system state. RMs account for the history dependence by maintaining an abstract state that explicitly tracks the QoS constraint violations over time. Our framework provides a principled, scalable approach to energy management for next-generation mobile networks under diverse traffic patterns and QoS requirements.

Recent work has explored optimizing image signal processing (ISP) pipelines for various tasks by composing predefined modules and adapting them to task-specific objectives. However, jointly optimizing module sequences and parameters remains challenging. Existing approaches rely on neural architecture search (NAS) or step-wise reinforcement learning (RL), but NAS suffers from a training-inference mismatch, while step-wise RL leads to unstable training and high computational overhead due to stage-wise decision-making. We propose POS-ISP, a sequence-level RL framework that formulates modular ISP optimization as a global sequence prediction problem. Our method predicts the entire module sequence and its parameters in a single forward pass and optimizes the pipeline using a terminal task reward, eliminating the need for intermediate supervision and redundant executions. Experiments across multiple downstream tasks show that POS-ISP improves task performance while reducing computational cost, highlighting sequence-level optimization as a stable and efficient paradigm for task-aware ISP. The project page is available at https://w1jyun.github.io/POS-ISP

In this paper, we study a vehicle-to-infrastructure (V2I) system where distributed base stations (BSs) acting as road-side units (RSUs) collect multimodal (wireless and visual) data from moving vehicles. We consider a decentralized rate maximization problem, where each RSU relies on its local observations to optimize its resources, while all RSUs must collaborate to guarantee favorable network performance. We recast this problem as a distributed multi-agent reinforcement learning (MARL) problem, by incorporating rotation symmetries in terms of vehicles' locations. To exploit these symmetries, we propose a novel self-supervised learning framework where each BS agent aligns the latent features of its multimodal observation to extract the positions of the vehicles in its local region. Equipped with this sensing data at each RSU, we train an equivariant policy network using a graph neural network (GNN) with message passing layers, such that each agent computes its policy locally, while all agents coordinate their policies via a signaling scheme that overcomes partial observability and guarantees the equivariance of the global policy. We present numerical results carried out in a simulation environment, where ray-tracing and computer graphics are used to collect wireless and visual data. Results show the generalizability of our self-supervised and multimodal sensing approach, achieving more than two-fold accuracy gains over baselines, and the efficiency of our equivariant MARL training, attaining more than 50% performance gains over standard approaches.

The problem of solving Markov decision processes under function approximation remains a fundamental challenge, even under linear function approximation settings. A key difficulty arises from a geometric mismatch: while the Bellman optimality operator is contractive in the Linfty-norm, commonly used objectives such as projected value iteration and Bellman residual minimization rely on L2-based formulations. To enable gradient-based optimization, we consider a soft formulation of Bellman residual minimization and extend it to a generalized weighted Lp -norm. We show that this formulation aligns the optimization objective with the contraction geometry of the Bellman operator as p increases, and derive corresponding performance error bounds. Our analysis provides a principled connection between residual minimization and Bellman contraction, leading to improved control of error propagation while remaining compatible with gradient-based optimization.

Value iteration-type methods have been extensively studied for computing a nearly optimal value function in reinforcement learning (RL). Under a generative sampling model, these methods can achieve sharper sample complexity than policy optimization approaches, particularly in their dependence on the discount factor. In practice, they are often employed for offline training. In this paper, we consider discounted Markov decision processes with state space S, action space A, discount factor $γ\in(0,1)$ and costs in $[0,1]$. We introduce a novel value optimization method, termed value mirror descent (VMD), which integrates mirror descent from convex optimization into the classical value iteration framework. In the deterministic setting with known transition kernels, we show that VMD converges linearly. For the stochastic setting with a generative model, we develop a stochastic variant, SVMD, which incorporates variance reduction commonly used in stochastic value iteration-type methods. For RL problems with general convex regularizers, SVMD attains a near-optimal sample complexity of $\tilde{O}(|S||A|(1-γ)^{-3}ε^{-2})$. Moreover, we establish that the Bregman divergence between the generated and optimal policies remains bounded throughout the iterations, even under the presence of model misspecification. This property is absent in existing stochastic value iteration-type methods but is important for enabling effective online (continual) learning following offline training. Under a strongly convex regularizer, SVMD achieves sample complexity of $\tilde{O}(|S||A|(1-γ)^{-5}ε^{-1})$, improving performance in the high-accuracy regime. Furthermore, we prove convergence of the generated policy to the optimal policy. Overall, the proposed method, its analysis, and the resulting guarantees, constitute new contributions to the RL and optimization literature.

Zero-shot unsupervised reinforcement learning (URL) offers a promising direction for building generalist agents capable of generalizing to unseen tasks without additional supervision. Among existing approaches, successor representations (SR) have emerged as a prominent paradigm due to their effectiveness in structured, low-dimensional settings. However, SR methods struggle to scale to high-dimensional visual environments. Through empirical analysis, we identify two key limitations of SR in visual URL: (1) SR objectives often lead to suboptimal representations that attend to dynamics-irrelevant regions, resulting in inaccurate successor measures and degraded task generalization; and (2) these flawed representations hinder SR policies from modeling multi-modal skill-conditioned action distributions and ensuring skill controllability. To address these limitations, we propose Saliency-Guided Representation with Consistency Policy Learning (SRCP), a novel framework that improves zero-shot generalization of SR methods in visual URL. SRCP decouples representation learning from successor training by introducing a saliency-guided dynamics task to capture dynamics-relevant representations, thereby improving successor measure and task generalization. Moreover, it integrates a fast-sampling consistency policy with URL-specific classifier-free guidance and tailored training objectives to improve skill-conditioned policy modeling and controllability. Extensive experiments on 16 tasks across 4 datasets from the ExORL benchmark demonstrate that SRCP achieves state-of-the-art zero-shot generalization in visual URL and is compatible with various SR methods.

Large language model (LLM) agents have demonstrated strong capabilities in complex interactive decision-making tasks. However, existing LLM agents typically rely on increasingly long interaction histories, resulting in high computational cost and limited scalability. In this paper, we propose STEP-HRL, a hierarchical reinforcement learning (HRL) framework that enables step-level learning by conditioning only on single-step transitions rather than full interaction histories. STEP-HRL structures tasks hierarchically, using completed subtasks to represent global progress of overall task. By introducing a local progress module, it also iteratively and selectively summarizes interaction history within each subtask to produce a compact summary of local progress. Together, these components yield augmented step-level transitions for both high-level and low-level policies. Experimental results on ScienceWorld and ALFWorld benchmarks consistently demonstrate that STEP-HRL substantially outperforms baselines in terms of performance and generalization while reducing token usage. Our code is available at https://github.com/TonyStark042/STEP-HRL.

A fundamental challenge in creative writing lies in reconciling the inherent tension between maintaining global coherence in long-form narratives and preserving local expressiveness in short-form texts. While long-context generation necessitates explicit macroscopic planning, short-form creativity often demands spontaneous, constraint-free expression. Existing alignment paradigms, however, typically employ static reward signals and rely heavily on high-quality supervised data, which is costly and difficult to scale. To address this, we propose UniCreative, a unified reference-free reinforcement learning framework. We first introduce AC-GenRM, an adaptive constraint-aware reward model that dynamically synthesizes query-specific criteria to provide fine-grained preference judgments. Leveraging these signals, we propose ACPO, a policy optimization algorithm that aligns models with human preferences across both content quality and structural paradigms without supervised fine-tuning and ground-truth references. Empirical results demonstrate that AC-GenRM aligns closely with expert evaluations, while ACPO significantly enhances performance across diverse writing tasks. Crucially, our analysis reveals an emergent meta-cognitive ability: the model learns to autonomously differentiate between tasks requiring rigorous planning and those favoring direct generation, validating the effectiveness of our direct alignment approach.

Unlike traditional homogeneous routing problems, the Heterogeneous Fleet Vehicle Routing Problem (HFVRP) involves heterogeneous fixed costs, variable travel costs, and capacity constraints, rendering solution quality highly sensitive to vehicle selection. Furthermore, real-world logistics applications often impose additional complex constraints, markedly increasing computational complexity. However, most existing Deep Reinforcement Learning (DRL)-based methods are restricted to homogeneous scenarios, leading to suboptimal performance when applied to HFVRP and its complex variants. To bridge this gap, we investigate HFVRP under complex constraints and develop a unified DRL framework capable of solving the problem across various variant settings. We introduce the Vehicle-as-Prompt (VaP) mechanism, which formulates the problem as a single-stage autoregressive decision process. Building on this, we propose VaP-CSMV, a framework featuring a cross-semantic encoder and a multi-view decoder that effectively addresses various problem variants and captures the complex mapping relationships between vehicle heterogeneity and customer node attributes. Extensive experimental results demonstrate that VaP-CSMV significantly outperforms existing state-of-the-art DRL-based neural solvers and achieves competitive solution quality compared to traditional heuristic solvers, while reducing inference time to mere seconds. Furthermore, the framework exhibits strong zero-shot generalization capabilities on large-scale and previously unseen problem variants, while ablation studies validate the vital contribution of each component.

Model-based reinforcement learning is attractive for sequential decision-making because it explicitly estimates reward and transition models and then supports planning through simulated rollouts. In offline settings with hidden confounding, however, models learned directly from observational data may be biased. This challenge is especially pronounced in partially observable systems, where latent factors may jointly affect actions, rewards, and future observations. Recent work has shown that policy evaluation in such confounded partially observable Markov decision processes (POMDPs) can be reduced to estimating reward-emission and observation-transition bridge functions satisfying conditional moment restrictions (CMRs). In this paper, we study the statistical estimation of these bridge functions. We formulate bridge learning as a CMR problem with nuisance objects given by a conditional mean embedding and a conditional density. We then develop a $K$-fold cross-fitted extension of the existing two-stage bridge estimator. The proposed procedure preserves the original bridge-based identification strategy while using the available data more efficiently than a single sample split. We also derive an oracle-comparator bound for the cross-fitted estimator and decompose the resulting error into a Stage I term induced by nuisance estimation and a Stage II term induced by empirical averaging.

This paper develops an end-to-end fuzzy encoder-decoder architecture for enhancing vision-based multi-modal deep spiking Q-networks in autonomous driving. The method addresses two core limitations of spiking reinforcement learning: information loss stemming from the conversion of dense visual inputs into sparse spike trains, and the limited representational capacity of spike-based value functions, which often yields weakly discriminative Q-value estimates. The encoder introduces trainable fuzzy membership functions to generate expressive, population-based spike representations, and the decoder uses a lightweight neural decoder to reconstruct continuous Q-values from spiking outputs. Experiments on the HighwayEnv benchmark show that the proposed architecture substantially improves decision-making accuracy and closes the performance gap between spiking and non-spiking multi-modal Q-networks. The results highlight the potential of this framework for efficient and real-time autonomous driving with spiking neural networks.

In this paper, we develop a stratification-based semantics for Signal Temporal Logic (STL) in which each atomic predicate is interpreted as a membership test in a stratified space. This perspective reveals a novel correspondence principle between stratification theory and STL, showing that most STL formulas can be viewed as inducing a stratification of space-time. The significance of this interpretation is twofold. First, it offers a fresh theoretical framework for analyzing the structure of the embedding space generated by deep reinforcement learning (DRL) and relates it to the geometry of the ambient decision space. Second, it provides a principled framework that both enables the reuse of existing high-dimensional analysis tools and motivates the creation of novel computational techniques. To ground the theory, we (1) illustrate the role of stratification theory in Minigrid games and (2) apply numerical techniques to the latent embeddings of a DRL agent playing such a game where the robustness of STL formulas is used as the reward. In the process, we propose computationally efficient signatures that, based on preliminary evidence, appear promising for uncovering the stratification structure of such embedding spaces.

Deep reinforcement learning (DRL) has shown remarkable performance on complex control problems in systems and networking, including adaptive video streaming, wireless resource management, and congestion control. For safe deployment, however, it is critical to reason about how agents behave across the range of system states they encounter in practice. Existing verification-based methods in this domain primarily focus on point properties, defined around fixed input states, which offer limited coverage and require substantial manual effort to identify relevant input-output pairs for analysis. In this paper, we study symbolic properties, that specify expected behavior over ranges of input states, for DRL agents in systems and networking. We present a generic formulation for symbolic properties, with monotonicity and robustness as concrete examples, and show how they can be analyzed using existing DNN verification engines. Our approach encodes symbolic properties as comparisons between related executions of the same policy and decomposes them into practically tractable sub-properties. These techniques serve as practical enablers for applying existing verification tools to symbolic analysis. Using our framework, diffRL, we conduct an extensive empirical study across three DRL-based control systems, adaptive video streaming, wireless resource management, and congestion control. Through these case studies, we analyze symbolic properties over broad input ranges, examine how property satisfaction evolves during training, study the impact of model size on verifiability, and compare multiple verification backends. Our results show that symbolic properties provide substantially broader coverage than point properties and can uncover non-obvious, operationally meaningful counterexamples, while also revealing practical solver trade-offs and limitations.

Training interpretable concept-based policies requires practitioners to manually select which human-understandable concepts an agent should reason with when making sequential decisions. This selection demands domain expertise, is time-consuming and costly, scales poorly with the number of candidates, and provides no performance guarantees. To overcome this limitation, we propose the first algorithms for principled automatic concept selection in sequential decision-making. Our key insight is that concept selection can be viewed through the lens of state abstraction: intuitively, a concept is decision-relevant if removing it would cause the agent to confuse states that require different actions. As a result, agents should rely on decision-relevant concepts; states with the same concept representation should share the same optimal action, which preserves the optimal decision structure of the original state space. This perspective leads to the Decision-Relevant Selection (DRS) algorithm, which selects a subset of concepts from a candidate set, along with performance bounds relating the selected concepts to the performance of the resulting policy. Empirically, DRS automatically recovers manually curated concept sets while matching or exceeding their performance, and improves the effectiveness of test-time concept interventions across reinforcement learning benchmarks and real-world healthcare environments.

In introductory presentations of policy gradients, one often derives the REINFORCE estimator using the full trajectory return and then states, by ``causality,'' that the full return may be replaced by the reward-to-go. Although this statement is correct, it is frequently presented at a level of rigor that leaves unclear where the past-reward terms disappear. This short paper isolates that step and gives a mathematically explicit derivation based on prefix trajectory distributions and the score-function identity. The resulting account does not change the estimator. Its contribution is conceptual: instead of presenting reward-to-go as a post hoc unbiased replacement for full return, it shows that reward-to-go arises directly once the objective is decomposed over prefix trajectories. In this formulation, the usual causality argument is recovered as a corollary of the derivation rather than as an additional heuristic principle.

This paper introduces Anticipatory Reinforcement Learning (ARL), a novel framework designed to bridge the gap between non-Markovian decision processes and classical reinforcement learning architectures, specifically under the constraint of a single observed trajectory. In environments characterised by jump-diffusions and structural breaks, traditional state-based methods often fail to capture the essential path-dependent geometry required for accurate foresight. We resolve this by lifting the state space into a signature-augmented manifold, where the history of the process is embedded as a dynamical coordinate. By utilising a self-consistent field approach, the agent maintains an anticipated proxy of the future path-law, allowing for a deterministic evaluation of expected returns. This transition from stochastic branching to a single-pass linear evaluation significantly reduces computational complexity and variance. We prove that this framework preserves fundamental contraction properties and ensures stable generalisation even in the presence of heavy-tailed noise. Our results demonstrate that by grounding reinforcement learning in the topological features of path-space, agents can achieve proactive risk management and superior policy stability in highly volatile, continuous-time environments.

Inference-time compute scaling has emerged as a powerful paradigm for improving language model performance on a wide range of tasks, but the question of how best to use the additional compute remains open. A popular approach is BoN sampling, where N candidate responses are generated, scored according to a reward model, and the highest-scoring response is selected. While this approach can improve performance, it is vulnerable to reward hacking, where performance degrades as N increases due to the selection of responses that exploit imperfections in the reward model instead of genuinely improving generation quality. Prior attempts to mitigate reward hacking, via stronger reward models or heavy-handed distributional regularization, either fail to fully address over-optimization or are too conservative to exploit additional compute. In this work, we explore the principle of pessimism in RL, which uses lower confidence bounds on value estimates to avoid OOD actions with uncertain reward estimates. Our approach, termed as caution, can be seen as the reverse of curiosity: where curiosity rewards prediction error as a signal of novelty, caution penalizes prediction error as a signal of distributional uncertainty. Practically, caution trains an error model on typical responses and uses its prediction error to lower reward estimates for atypical ones. Our extensive empirical evaluation demonstrates that caution is a simple, computationally efficient approach that substantially mitigates reward hacking in BoN sampling. We also provide a theoretical analysis in a simplified linear setting, which shows that caution provably improves over the standard BoN approach. Together, our results not only establish caution as a practical solution to reward hacking, but also provide evidence that curiosity-based approaches can be a general OOD detection technique in LLM settings.

Researchers and practitioners are increasingly considering reinforcement learning to optimize decisions in complex domains like robotics and healthcare. To date, these efforts have largely utilized expectation-based learning. However, relying on expectation-focused objectives may be insufficient for making consistent decisions in highly uncertain situations involving multiple heterogeneous groups. While distributional reinforcement learning algorithms have been introduced to model the full distributions of outcomes, they can yield large discrepancies in realized benefits among comparable agents. This challenge is particularly acute in healthcare settings, where physicians (controllers) must manage multiple patients (subordinate agents) with uncertain disease progression and heterogeneous treatment responses. We propose a Boosted Distributional Reinforcement Learning (BDRL) algorithm that optimizes agent-specific outcome distributions while enforcing comparability among similar agents and analyze its convergence. To further stabilize learning, we incorporate a post-update projection step formulated as a constrained convex optimization problem, which efficiently aligns individual outcomes with a high-performing reference within a specified tolerance. We apply our algorithm to manage hypertension in a large subset of the US adult population by categorizing individuals into cardiovascular disease risk groups. Our approach modifies treatment plans for median and vulnerable patients by mimicking the behavior of high-performing references in each risk group. Furthermore, we find that BDRL improves the number and consistency of quality-adjusted life years compared with reinforcement learning baselines.

Despite the sustained popularity of Q-learning as a practical tool for policy determination, a majority of relevant theoretical literature deals with either constant ($η_{t}\equiv η$) or polynomially decaying ($η_{t} = ηt^{-α}$) learning schedules. However, it is well known that these choices suffer from either persistent bias or prohibitively slow convergence. In contrast, the recently proposed linear decay to zero (LD2Z: $η_{t,n}=η(1-t/n)$) schedule has shown appreciable empirical performance, but its theoretical and statistical properties remain largely unexplored, especially in the Q-learning setting. We address this gap in the literature by first considering a general class of power-law decay to zero (PD2Z-$ν$: $η_{t,n}=η(1-t/n)^ν$). Proceeding step-by-step, we present a sharp non-asymptotic error bound for Q-learning with PD2Z-$ν$ schedule, which then is used to derive a central limit theory for a new tail Polyak-Ruppert averaging estimator. Finally, we also provide a novel time-uniform Gaussian approximation (also known as strong invariance principle) for the partial sum process of Q-learning iterates, which facilitates bootstrap-based inference. All our theoretical results are complemented by extensive numerical experiments. Beyond being new theoretical and statistical contributions to the Q-learning literature, our results definitively establish that LD2Z and in general PD2Z-$ν$ achieve a best-of-both-worlds property: they inherit the rapid decay from initialization (characteristic of constant step-sizes) while retaining the asymptotic convergence guarantees (characteristic of polynomially decaying schedules). This dual advantage explains the empirical success of LD2Z while providing practical guidelines for inference through our results.

Limited work has examined the strategic behaviors of relational networked learning agents under social dilemmas, and has overlooked the intricate social dynamics of complex systems. We address the challenge with Socio-Relational Intrinsic Motivation (SRIM), which endows agents with diverse preferences over sub-graphical social structures in order to study the impact of agents' personal preferences over their sub-graphical relations on their strategic decision-making under sequential social dilemmas. Our results in the Harvest and Cleanup environments demonstrate that preferences over different subgraph structures (degree-, clique-, and critical connection-based) lead to distinct variations in agents' reward gathering and strategic behavior: individual aggressiveness in Harvest and individual contribution effort in Cleanup. Moreover, agents with different subgraphical structural positions consistently exhibit similar strategic behavioral shifts. Our proposed BCI metric captures structural variation within the population, and the relative ordering of BCI across social preferences is consistent in Harvest and Cleanup games for the same topology, suggesting the subgraphical structural impact is robust across environments. These results provide a new lens for examining agents' behavior in social dilemmas and insight for designing effective multi-agent ecosystems composed of heterogeneous social agents.

Reinforcement learning in real-world systems often involves delayed feedback, which breaks the Markov assumption and impedes both learning and control. Canonical augmentation-based approaches cause state-space explosion, which imposes a severe sample-complexity burden. Despite recent progress, state-of-the-art augmentation-based baselines either mainly alleviate the burden on the critic or rely on non-unified treatments for the actor and critic. In this study, we propose delayed homomorphic reinforcement learning (DHRL), a framework grounded in MDP homomorphisms that defines a belief-equivalence relation over the augmented state space to collapse control-redundant augmented states. In principle, this yields exact abstraction under deterministic dynamics and approximate abstraction under stochastic dynamics, enabling both the actor and critic to benefit from a structured abstraction mechanism. In finite domains, exact abstraction preserves optimality and recovers the delay-free sample-complexity order, whereas approximate abstraction admits a value-loss bound on the resulting policy. For continuous domains, we introduce deep delayed homomorphic policy gradient (D$^2$HPG), a deep actor-critic instantiation of the DHRL framework. Experiments on continuous-control tasks in MuJoCo show that D$^2$HPG outperforms strong augmentation-based baselines.

Training tool-calling agents with reinforcement learning on multi-turn tasks remains challenging due to sparse outcome rewards and difficult credit assignment across conversation turns. We present the first application of MT-GRPO (Multi-Turn Group Relative Policy Optimization) combined with GTPO (Generalized Token-level Policy Optimization) for training a tool-calling agent on realistic customer service tasks with an LLM-based user simulator. Through systematic analysis of training rollouts, we discover that naively designed dense per-turn rewards degrade performance by up to 14 percentage points due to misalignment between reward discriminativeness and advantage direction. We introduce Iterative Reward Calibration, a methodology for designing per-turn rewards using empirical discriminative analysis of rollout data, and show that our GTPO hybrid advantage formulation eliminates the advantage misalignment problem. Applied to the Tau-Bench airline benchmark, our approach improves Qwen3.5-4B from 63.8 percent to 66.7 percent (+2.9pp) and Qwen3-30B-A3B from 58.0 percent to 69.5 percent (+11.5pp) -- with the trained 4B model exceeding GPT-4.1 (49.4 percent) and GPT-4o (42.8 percent) despite being 50 times smaller, and the 30.5B MoE model approaching Claude Sonnet 4.5 (70.0 percent). To our knowledge, these are the first published RL training results on Tau-Bench. We release our code, reward calibration analysis, and training recipes.

Quadrupedal locomotion over complex terrain has been a long-standing research topic in robotics. While recent reinforcement learning-based locomotion methods improve generalizability and foot-placement precision, they rely on implicit inference of foot positions from joint angles, lacking the explicit precision and stability guarantees of optimization-based approaches. To address this, we introduce a foot position map integrated into the heightmap, and a dynamic locomotion-stability reward within an attention-based framework to achieve locomotion on complex terrain. We validate our method extensively on terrains seen during training as well as out-of-domain (OOD) terrains. Our results demonstrate that the proposed method enables precise and stable movement, resulting in improved locomotion success rates on both in-domain and OOD terrains.

Reinforcement learning (RL) agents often struggle with out-of-distribution (OOD) scenarios, leading to high uncertainty and random behavior. While language models (LMs) contain valuable world knowledge, larger ones incur high computational costs, hindering real-time use, and exhibit limitations in autonomous planning. We introduce Adaptive Safety through Knowledge (ASK), which combines smaller LMs with trained RL policies to enhance OOD generalization without retraining. ASK employs Monte Carlo Dropout to assess uncertainty and queries the LM for action suggestions only when uncertainty exceeds a set threshold. This selective use preserves the efficiency of existing policies while leveraging the language model's reasoning in uncertain situations. In experiments on the FrozenLake environment, ASK shows no improvement in-domain, but demonstrates robust navigation in transfer tasks, achieving a reward of 0.95. Our findings indicate that effective neuro-symbolic integration requires careful orchestration rather than simple combination, highlighting the need for sufficient model scale and effective hybridization mechanisms for successful OOD generalization.

Rerankers play a pivotal role in refining retrieval results for Retrieval-Augmented Generation. However, current reranking models are typically optimized on static human annotated relevance labels in isolation, decoupled from the downstream generation process. This isolation leads to a fundamental misalignment: documents identified as topically relevant by information retrieval metrics often fail to provide the actual utility required by the LLM for precise answer generation. To bridge this gap, we introduce ReRanking Preference Optimization (RRPO), a reinforcement learning framework that directly aligns reranking with the LLM's generation quality. By formulating reranking as a sequential decision-making process, RRPO optimizes for context utility using LLM feedback, thereby eliminating the need for expensive human annotations. To ensure training stability, we further introduce a reference-anchored deterministic baseline. Extensive experiments on knowledge-intensive benchmarks demonstrate that RRPO significantly outperforms strong baselines, including the powerful list-wise reranker RankZephyr. Further analysis highlights the versatility of our framework: it generalizes seamlessly to diverse readers (e.g., GPT-4o), integrates orthogonally with query expansion modules like Query2Doc, and remains robust even when trained with noisy supervisors.