Reinforcement Learning

Also known as RL

Follow this topic Ask about this topic

Reinforcement learning studies how an agent can choose actions over time to earn rewards, using experience to improve its decisions. Common approaches include value-based methods such as Q-learning, policy-gradient methods such as PPO, and model-based methods that learn or use a model of the environment; offline and hierarchical methods address learning from fixed data and long tasks. Applications range from Atari and MuJoCo benchmarks to robot control, traffic, games, and language-model agents that use tools or solve multi-step problems. A central challenge is making learning efficient and reliable when rewards are sparse or imperfect, environments change, and success in simulation or benchmarks may not transfer safely to deployment.

Latest papers

1,742 papers237 papers in the last four weeks, up 46% on the four weeks before. 1.4% of all new papers.

Jan 20, 2026

KAGE-Bench: Fast Known-Axis Visual Generalization Evaluation for Reinforcement Learning

Pixel-based reinforcement learning agents often fail under purely visual distribution shift even when latent dynamics and rewards are unchanged, but existing benchmarks entangle multiple sources of shift and hinder systematic analysis. We introduce KAGE-Env, a JAX-native 2D platformer that factorizes the observation process into independently controllable visual axes while keeping the underlying control problem fixed. By construction, varying a visual axis affects performance only through the induced state-conditional action distribution of a pixel policy, providing a clean abstraction for visual generalization. Building on this environment, we define KAGE-Bench, a benchmark of six known-axis suites comprising 34 train-evaluation configuration pairs that isolate individual visual shifts. Using a standard PPO-CNN baseline, we observe strong axis-dependent failures, with background and photometric shifts often collapsing success, while agent-appearance shifts are comparatively benign. Several shifts preserve forward motion while breaking task completion, showing that return alone can obscure generalization failures. Finally, the fully vectorized JAX implementation enables up to 33M environment steps per second on a single GPU, enabling fast and reproducible sweeps over visual factors. Code: https://avanturist322.github.io/KAGEBench/.

Egor Cherepanov, Daniil Zelezetsky, Alexey K. Kovalev +1

1AXXX, Moscow, Russia · 2MIRAI, Moscow, Russia

RL Benchmarks Reinforcement Learning Domain Generalization Visual RL Generalization in RL

Jan 15, 2026

Reinforcement Learning with Multi-Step Lookahead Information Via Adaptive Batching

We study tabular finite-horizon reinforcement learning problems with multiple steps of lookahead information. Before acting, the learner observes

\ell

steps of future transition and reward realizations: the exact state the agent would reach and the rewards it would collect under any possible course of action. While it has been shown that such information can drastically boost the value, finding the optimal policy is NP-hard, and it is common to apply one of two tractable heuristics: processing the lookahead in chunks of predefined sizes ('fixed batching policies'), and model predictive control. We first illustrate the problems with these two approaches and propose utilizing the lookahead in adaptive (state-dependent) batches; we refer to such policies as adaptive batching policies (ABPs). We derive the optimal Bellman equations for these strategies and design an optimistic regret-minimizing algorithm that enables learning the optimal ABP when interacting with unknown environments. Our regret bounds are order-optimal up to a potential factor of the lookahead horizon

\ell

, which can usually be considered a small constant.

Nadav Merlis

The Faculty of Data and Decision Sciences, Technion – Israel Institute of Technology, Israel

Reinforcement Learning Regret Minimization in RL

Jan 9, 2026

MemBuilder: Reinforcing LLMs for Long-Term Memory Construction via Attributed Dense Rewards

Maintaining consistency in long-term dialogues remains a fundamental challenge for LLMs, as standard retrieval mechanisms often fail to capture the temporal evolution of historical states. While memory-augmented frameworks offer a structured alternative, current systems rely on static prompting of closed-source models or suffer from ineffective training paradigms with sparse rewards. We introduce MemBuilder, a reinforcement learning framework that trains models to orchestrate multi-dimensional memory construction with attributed dense rewards. MemBuilder addresses two key challenges: (1) Sparse Trajectory-Level Rewards: we employ synthetic session-level question generation to provide dense intermediate rewards across extended trajectories; and (2) Multi-Dimensional Memory Attribution: we introduce contribution-aware gradient weighting that scales policy updates based on each component's downstream impact. Experimental results show that MemBuilder enables a 4B-parameter model to outperform state-of-the-art closed-source baselines, exhibiting strong generalization across long-term dialogue benchmarks.

Zhiyu Shen, Ziming Wu, Fuming Lai +2

School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · Tencent Inc.

Reinforcement Learning Memory-Augmented Language Models Memorization in Language Models Persistent Memory for Language Models LLM Memory Long-Term Conversational Memory+1

Jan 8, 2026

RL-AWB: Deep Reinforcement Learning for Auto White Balance Correction in Low-Light Night-time Scenes

Nighttime color constancy still remains a challenging problem in computational photography due to low-light noise and complex illumination conditions. We present RL-AWB, a novel framework combining statistical methods with deep reinforcement learning for nighttime white balance. Our method begins with a statistical algorithm tailored for nighttime scenes, integrating salient gray pixel detection with novel illuminant estimation. Building on this foundation, we develop the first deep reinforcement learning approach for color constancy that leverages the statistical algorithm as its core, mimicking professional AWB tuning experts by dynamically determining image-specific parameters at inference time, without requiring ground-truth illuminants or reference images. To further facilitate cross-sensor evaluation, we introduce the first multi-sensor nighttime dataset. Experiment results demonstrate that our method achieves strong generalization capability across low-light and well-illuminated images. Project page: https://ntuneillee.github.io/research/rl-awb/

Yuan-Kang Lee, Kuan-Lin Chen, Chia-Che Chang +1

MediaTek Inc., Taiwan · National Taiwan University, Taiwan · National Yang Ming Chiao Tung University, Taiwan

Image Enhancement Reinforcement Learning Color Constancy Low-Light Image Enhancement

Jan 7, 2026

Survival Dynamics of Neural and Programmatic Policies in Evolutionary Reinforcement Learning

In evolutionary reinforcement learning tasks (ERL), agent policies are often encoded as small artificial neural networks (NERL). Such representations lack explicit modular structure, limiting behavioral interpretation. We investigate whether programmatic policies (PERL), implemented as soft, differentiable decision lists (SDDL), can match the performance of NERL. To support reproducible evaluation, we provide the first fully specified and open-source reimplementation of the classic 1992 Artificial Life (ALife) ERL testbed. We conduct a rigorous survival analysis across 4000 independent trials utilizing Kaplan-Meier curves and Restricted Mean Survival Time (RMST) metrics absent in the original study. We find a statistically significant difference in survival probability between PERL and NERL. PERL agents survive on average 201.69 steps longer than NERL agents. Moreover, SDDL agents using learning alone (no evolution) survive on average 73.67 steps longer than neural agents using both learning and evaluation. These results demonstrate that programmatic policies can exceed the survival performance of neural policies in ALife.

Anton Roupassov-Ruiz, Yiyang Zuo

Department of Computing Science University of Alberta Edmonton, Alberta, T6G 2E8, Canada

Reinforcement Learning Artificial Life Policy Optimization Evolutionary RL

Dec 30, 2025

Soft Fitted Q-Iteration without Bellman Completeness: Occupancy Reweighting and Temperature Annealing

Fitted

Q

-iteration (FQI) is a standard regression-based method for optimal control in offline reinforcement learning, but its stability under function approximation often relies on Bellman completeness, which requires Bellman images of the fitted class to remain in the class. We study Kullback--Leibler (KL)-regularized, or soft, FQI relative to a fixed reference policy without this assumption. Our key insight is that soft control locally inherits the contraction of policy evaluation in a discounted-occupancy norm. At the soft-optimal fixed point, the linearization of the soft Bellman operator is exactly the Bellman operator for the soft-optimal policy, which contracts in its discounted-occupancy norm; projection in the same norm preserves this contraction. Standard soft FQI instead projects under the offline state-action distribution and need not preserve this property. Motivated by this observation, we propose occupancy-reweighted soft FQI, which retains standard Bellman targets and least-squares updates while reweighting regressions by discounted-occupancy ratios induced by the current soft policy. Under

Q

-function realizability and local regularity, we establish local contraction and finite-sample convergence with estimated ratios, without Bellman completeness. We then use temperature annealing to convert the local result into global convergence from arbitrary initialization: sufficiently high temperature provides a globally contractive starting regime, while gradual cooling connects successive local contraction regions to any prescribed positive target temperature. Under an action-gap margin condition, switching at a fixed positive temperature to hard FQI with refreshed occupancy weights also yields population and finite-sample convergence to the unregularized optimum.

Lars van der Laan, Nathan Kallus

Department of Statistics, University of Washington · Netflix · Cornell University

Reinforcement Learning KL-Regularized RL Q-Learning Offline RL Maximum Entropy RL Policy Learning+1

Dec 28, 2025

Trust Region Masking for Long-Horizon LLM Reinforcement Learning

Policy gradient methods for Large Language Models optimize a policy

π_θ

via a surrogate objective computed from samples of a rollout policy

π_{\text{roll}}

. However, modern LLM-RL pipelines suffer from unavoidable implementation divergences -- backend discrepancies, Mixture-of-Experts routing discontinuities, and distributed training staleness -- causing off-policy mismatch (

π_{\text{roll}} \neq π_θ

) and approximation errors between the surrogate and the true objective. We demonstrate that classical trust region bounds on this error scale as

O(T^2)

with sequence length

T

, rendering them vacuous for long-horizon tasks. To address this, we derive a family of bounds -- both KL-based and TV-based -- including a Pinsker-Marginal bound (

O(T^{3/2})

), a Mixed bound (

O(T)

), and an Adaptive bound that strictly generalizes the Pinsker-Marginal bound via per-position importance-ratio decomposition. Taking the minimum over all bounds yields the tightest known guarantee across all divergence regimes. Crucially, all bounds depend on the maximum token-level divergence

D_{\mathrm{KL}}^{\mathrm{tok,max}}

(or

D_{\mathrm{TV}}^{\mathrm{tok,max}}

), a sequence-level quantity that cannot be controlled by token-independent methods like PPO clipping. We propose Trust Region Masking (TRM), which masks entire sequences violating the trust region, enabling the first non-vacuous monotonic improvement guarantees for long-horizon LLM-RL.

Yingru Li, Jiacai Liu, Jiawei Xu +4

Fudan University · The Chinese University of Hong Kong, Shenzhen

RL for Language Models Reinforcement Learning Long-Horizon RL Policy Gradient Methods Trust Region Policy Optimization Policy Learning+1

Dec 24, 2025

Meta-RL with Bayesian Linear Task Models

Deep Bayesian reinforcement learning adapts to unseen tasks by inferring latent transition and reward models, but existing methods typically rely on variational posteriors and evidence lower bounds, introducing approximation error and unstable task representations. We introduce GLiBRL, a deep Bayesian RL framework that combines generalised linear task models with learnable non-linear basis functions. GLiBRL features conjugate Bayesian inference, yielding exact, sequential posterior updates over task parameters and model noise, together with a closed-form marginal likelihood that eliminates variational inference. The update is naturally permutation-invariant, allowing GLiBRL to integrate with both off- and on-policy algorithms. GLiBRL also learns task representation admitting an exact kernel identity, relating distances between task representations to kernel discrepancies over the task contexts. Compared against eight representative or recent meta reinforcement learning methods, GLiBRL achieves the highest aggregate zero-shot test performance on both the MuJoCo locomotion and MetaWorld manipulation benchmarks.

Jingyang You, Hanna Kurniawati

School of Computing Australian National University

Representation Learning Reinforcement Learning Bayesian RL Meta-Reinforcement Learning Generalization in RL Model-Based RL

Dec 19, 2025

The HydroGym Reinforcement Learning Platform for Fluid Dynamics

Modeling and controlling fluids is critical across science and engineering. Effective flow control can increase lift, reduce drag, enhance mixing, and attenuate noise, potentially unlocking new technologies. Yet controlling fluids is hard: the dynamics are high-dimensional, nonlinear, and multiscale. While reinforcement learning (RL) has recently succeeded in robotics and protein folding through shared benchmarks, fluid dynamics has resisted such progress: each controller is typically tuned to a single geometry and operating point, making results hard to accumulate, transfer, and compare. We introduce HydroGym, a solver-independent RL platform for flow control, and show that standardized infrastructure unlocks transferable control intelligence across flow regimes. HydroGym provides 61+ validated environments spanning laminar to turbulent flows, with systematic Reynolds number progressions up to Re=400,000 and Mach number variations in 2D and 3D. It supports diverse backends, including finite-volume, spectral-element, finite-element, lattice-Boltzmann, and fully differentiable solvers for gradient-enhanced optimization. Across environments, RL agents consistently discover robust control principles, such as boundary-layer manipulation, acoustic-feedback disruption, and wake reorganization, yielding drag reductions exceeding 90% in canonical configurations. Critically, we demonstrate zero-shot transfer: agents trained only on a simplified channel flow achieve 38% friction-drag reduction on an unseen 3D wing section at chord Reynolds number Re=200,000 reducing exploration costs by four orders of magnitude versus direct on-wing optimization. This suggests RL agents uncover essential physics rather than configuration-specific patterns, pointing toward generalizable control. HydroGym offers extensible, scalable community infrastructure for fluid dynamics, machine learning, and control research.

Christian Lagemann, Sajeda Mokbel, Miro Gondrum +18

Department of Mechanical Engineering, University of Washington, Seattle, WA 98195, United States · AI Institute in Dynamic Systems, University of Washington, Seattle, WA 98195, United States · Chair of Fluid Mechanics and Institute of Aerodynamics, RWTH Aachen University, Aachen, Germany +9

RL Control RL Benchmarks Reinforcement Learning Computational Fluid Dynamics

Dec 17, 2025

Adaptive Partitioning and Learning for Stochastic Control of Diffusion Processes

We study reinforcement learning for controlled diffusion processes with unbounded continuous state spaces, bounded continuous actions, and polynomially growing rewards: settings that arise naturally in finance, economics, and operations research. To overcome the challenges of continuous and high-dimensional domains, we introduce a model-based algorithm that adaptively partitions the joint state-action space. The algorithm maintains estimators of drift, volatility, and rewards within each partition, refining the discretization whenever estimation bias exceeds statistical confidence. This adaptive scheme balances exploration and approximation, enabling efficient learning in unbounded domains. Our analysis establishes regret bounds that depend on the problem horizon, state dimension, reward growth order, and a newly defined notion of zooming dimension tailored to unbounded diffusion processes. The bounds recover existing results for bounded settings as a special case, while extending theoretical guarantees to a broader class of diffusion-type problems. Finally, we validate the effectiveness of our approach through numerical experiments, including applications to high-dimensional problems such as multi-asset mean-variance portfolio selection.

Hanqing Jin, Renyuan Xu, Yanzhao Yang

Mathematical Institute, University of Oxford · Department of Management Science and Engineering, Stanford University

RL Control Reinforcement Learning Stochastic Optimal Control Model-Based RL

Dec 15, 2025

Dual-Evolving Reinforcement Learning for Automated Reward Design

Crafting effective reward signals remains a central challenge in reinforcement learning (RL) for complex reasoning. Existing methods for automated reward optimization search with a fixed proposer which never learns from its costly trials. We present Dual-Evolving Reinforcement Learning (DERL), a dual-training framework for autonomous discovery of reward structures. An outer-loop proposer composes a reward from task primitives, and an inner policy trains with this reward. The proposer itself learns by policy gradient from the validation performance of the inner policy. Thus, DERL utilizes past-trial feedback to shape its reward-proposal distribution, an empirical meta-gradient of task success. We validate DERL on embodied (ALFWorld), scientific (ScienceWorld), and mathematical (GSM8K, MATH) reasoning. On multi-turn agent benchmarks, DERL outperforms strong baselines, significantly on out-of-distribution tasks. In mathematics, DERL improves on the outcome reward where directly added primitives hurt. Further analyses under matched compute show that our proposer effectively captures the meta-gradient to discover useful reward signals.

Sitao Cheng, Tianle Li, Xuhan Huang +2

University of Waterloo · The University of Hong Kong · The Chinese University of Hong Kong, Shenzhen +1

Reward Design for RL Reinforcement Learning RL for Language Model Reasoning Meta-Reinforcement Learning

Dec 12, 2025

Symmetry-Aware Steering of Equivariant Diffusion Policies: Benefits and Limits

Equivariant diffusion policies (EDPs) combine the expressivity of diffusion models with the generalization and sample efficiency afforded by geometric symmetries. While steering these policies with reinforcement learning (RL) offers a promising fine-tuning mechanism beyond demonstration data, directly applying standard (non-equivariant) RL ignores the symmetries that EDPs are designed to exploit, leading to sample inefficiency and instability. We theoretically establish that the diffusion process of an EDP is equivariant, inducing a group-invariant latent-noise MDP which is well-suited for equivariant diffusion steering. Building on this, we introduce a principled symmetry-aware steering framework and compare standard, equivariant, and approximately equivariant RL strategies across tasks with varying degrees of symmetry. While we identify the practical boundaries of strict equivariance under symmetry breaking, exploiting symmetry during the steering yields substantial benefits---enhancing sample efficiency, preventing value divergence, and achieving strong policy improvements even when EDPs are trained from extremely limited demonstrations.

Minwoo Park, Junwoo Chang, Jongeun Choi +1

Yonsei University, Seoul 03722, Republic of Korea · University of California, Berkeley, CA 94720, USA

Diffusion Policy RL for Diffusion Models Equivariant Policy Learning Reinforcement Learning Sample-Efficient RL

Dec 11, 2025

UBCL: A Reinforcement Learning Framework for Controllable and Diverse Player Behaviors

This paper introduces a reinforcement learning framework that enables controllable and diverse player behaviors without relying on human gameplay data. Existing approaches often require large-scale player trajectories, train separate models for different player types, or provide no direct mapping between interpretable behavioral parameters and the learned policy, limiting their scalability and controllability. We define player behavior in an N-dimensional continuous space and uniformly sample target behavior vectors from a region that encompasses the subset representing real human styles. During training, each agent receives both its current and target behavior vectors as input, and the reward is based on the normalized reduction in distance between them. This allows the policy to learn how actions influence behavioral statistics, enabling smooth control over attributes such as aggressiveness, mobility, and cooperativeness. A single PPO-based multi-agent policy can reproduce new or unseen play styles without retraining. Experiments conducted in a custom multi-player Unity game show that the proposed framework produces significantly greater behavioral diversity than a win-only baseline and reliably matches specified behavior vectors across diverse targets. The method offers a scalable solution for automated playtesting, game balancing, human-like behavior simulation, and replacing disconnected players in online games.

Atahan Cilan, Atay Özgövde

Turkish Aerospace, Istanbul, Turkey · Department of Computer Engineering, Boğaziçi University, Istanbul, Turkey

Reinforcement Learning Human Behavior Simulation Goal-Conditioned RL Multi-Agent Reinforcement Learning

Dec 11, 2025

Achieving Olympiad-Level Geometry Large Language Model Agent via Complexity Boosting Reinforcement Learning

Large language model (LLM) agents exhibit strong mathematical problem-solving abilities and can even solve International Mathematical Olympiad (IMO) level problems with the assistance of formal proof systems. However, due to weak heuristics for auxiliary constructions, AI for geometry problem solving remains dominated by expert models such as AlphaGeometry 2, which rely heavily on large-scale data synthesis and search for both training and evaluation. In this work, we make the first attempt to build a medalist-level LLM agent for geometry and present InternGeometry. InternGeometry overcomes the heuristic limitations in geometry by iteratively proposing propositions and auxiliary constructions, verifying them with a symbolic engine, and reflecting on the engine's feedback to guide subsequent proposals. A dynamic memory mechanism enables InternGeometry to conduct more than two hundred interactions with the symbolic engine per problem. To further accelerate learning, we introduce Complexity-Boosting Reinforcement Learning (CBRL), which gradually increases the complexity of synthesized problems across training stages. Built on InternThinker-32B, InternGeometry solves 44 of 50 IMO geometry problems (2000-2024), exceeding the average gold medalist score (40.9), using only 13K training examples, just 0.004% of the data used by AlphaGeometry 2, demonstrating the potential of LLM agents on expert-level geometry tasks. InternGeometry can also propose novel auxiliary constructions for IMO problems that do not appear in human solutions.

Haiteng Zhao, Junhao Shen, Yiming Zhang +7

1Shanghai AI Laboratory · 2Shanghai Jiao Tong University · 3Peking University +2

LLM Agents Reinforcement Learning Neuro-Symbolic Reasoning

Dec 7, 2025

Statistical analysis of Inverse Entropy-regularized Reinforcement Learning

Inverse reinforcement learning aims to infer the reward function that explains expert behavior observed through trajectories of state--action pairs. A long-standing difficulty in classical IRL is the non-uniqueness of the recovered reward: many reward functions can induce the same optimal policy, rendering the inverse problem ill-posed. In this paper, we develop a statistical framework for Inverse Entropy-regularized Reinforcement Learning that resolves this ambiguity by combining entropy regularization with a least-squares reconstruction of the reward from the soft Bellman residual. This combination yields a unique and well-defined so-called least-squares reward consistent with the expert policy. We model the expert demonstrations as a Markov chain with the invariant distribution defined by an unknown expert policy

π^\star

and estimate the policy by a penalized maximum-likelihood procedure over a class of conditional distributions on the action space. We establish high-probability bounds for the excess Kullback--Leibler divergence between the estimated policy and the expert policy, accounting for statistical complexity through covering numbers of the policy class. These results lead to non-asymptotic minimax optimal convergence rates for the least-squares reward function, revealing the interplay between smoothing (entropy regularization), model complexity, and sample size. Our analysis bridges the gap between behavior cloning, inverse reinforcement learning, and modern statistical learning theory.

Denis Belomestny, Alexey Naumov, Artemy Rubtsov +1

Reward Modeling Reinforcement Learning

Dec 2, 2025

CUDA-L2: Surpassing cuBLAS Performance for Matrix Multiplication through Reinforcement Learning

In this paper, we propose CUDA-L2, a system that combines large language models (LLMs) and reinforcement learning (RL) to automatically optimize Half-precision General Matrix Multiply (HGEMM) CUDA kernels. Using CUDA execution speed as the RL reward, CUDA-L2 automatically optimizes HGEMM kernels across 1,000 configurations. CUDA-L2 systematically outperforms major matmul baselines to date, from the widely-used torch.matmul to state-of-the-art Nvidia's closed-source libraries, i.e., cuBLAS, cuBLASLt. In offline mode, where kernels are executed consecutively without time intervals, CUDA-L2 yields +22.0% over torch.matmul on average; +19.2% over cuBLAS using the optimal layout configuration (normal-normal NN and transposed-normal TN); +16.8% over cuBLASLt-heuristic, which queries cuBLASLt library and selects the algorithm based on the heuristic's suggestion; and +11.4% over the most competitive cuBLASLt-AutoTuning model, which selects the fastest algorithm from up to 100 candidates from cuBLASLt's suggestions. In server mode, where kernels are executed at random intervals simulating real-time inference, the speedups further increase to +28.7%, +26.0%, +22.4%, and +15.9% for torch.matmul, cuBLAS, cuBLASLt-heuristic, and cuBLASLt-AutoTuning respectively. CUDA-L2 shows that even the most performance-critical, heavily-optimized kernels like HGEMM can be improved through LLM-guided RL automation by systematically exploring configuration spaces at scales impractical for humans. Project and code can be found at github.com/deepreinforce-ai/CUDA-L2

Songqiao Su, Xiaoya Li, Albert Wang +3

DeepReinforce Team

Reinforcement Learning GPU Kernel Optimization GPU Acceleration High-Performance Computing Large Language Model-Guided Optimization GPU Kernel Generation

Nov 26, 2025

Iterative Policy Optimization: Batch-Online Multi-Turn RL via Single-Turn RL

Practical LLM agents often operate over multi-turn conversations where success is determined only after the full interaction ends. Most multi-turn RL methods train via on-policy rollouts, but unlike in single-turn RL, the policy cannot produce a trajectory alone, since an external environment must respond after each agent turn. For conversational agents, this environment is a user, but real users are generally unavailable inside the training loop and simulated users are difficult to build faithfully. Separately, real-world deployment is rarely fully online or fully offline. The common production pattern is called batch online, where the current policy is deployed to collect a batch of interaction data, then retrained on that batch and redeployed. We show that this batch-online setting creates an opportunity for applying the classical approximate policy iteration algorithm. Our central observation is that running standard token-level, single-turn GRPO with a learned turn-level

Q

-function as the reward model is a policy improvement step for the multi-turn problem. Building on this, we present Iterative Policy Optimization (IterPO), which alternates between fitting

Q^π

from logged Monte Carlo returns via standard reward modeling (policy evaluation) and running single-turn GRPO against

Q^π

(policy improvement). During policy improvement, the learned

Q^π

scores candidate responses by their expected downstream return, avoiding the need for simulator rollouts of the remaining conversation. Overall, this allows us to do principled multi-turn RL using only single-turn RL methods without an interactive environment or user simulator inside the training loop. We demonstrate the effectiveness of IterPO on six multi-turn negotiation environments.

Daniel R. Jiang, Ankur Samanta, Yukai Yang +3

Meta · Columbia University · FAIR at Meta

Reinforcement Learning Policy Iteration Multi-Turn RL Policy Learning RL for LLM Agents

Nov 25, 2025

Attention Trajectories as a Diagnostic Axis for Deep Reinforcement Learning

The emergence and evolution of feature reliance in deep reinforcement learning agents remain poorly understood. Here, we introduce a methodological framework for analyzing the learning process through quantitative analysis of saliency maps. This approach aggregates saliency information at the object and modality level into hierarchical attention profiles, quantifying how agents allocate attention over time, thereby forming attention trajectories throughout training. These profiles are then compared across controlled conditions, connected to behavioral measurements and reproduced with different saliency methods to assess the robustness of the findings. Applied to Atari 2600 benchmarks, custom Pong environments, and biomechanical user simulations in visuomotor tasks, this framework uncovers algorithm-specific attention biases, diagnosed unintended reward-driven strategies, and overfitting to redundant sensory channels. These patterns correspond to measurable behavioral differences, demonstrating empirical links between attention profiles, learning dynamics, and agent behavior. The results establish attention trajectories as a promising diagnostic axis for tracing how feature reliance develops during training and for identifying biases and vulnerabilities invisible to performance metrics alone.

Charlotte Beylier, Hannah Selder, Arthur Fleig +2

Max Planck Institute for Human Cognitive and Brain Sciences · Center for Scalable Data Analytics and Artificial Intelligence (ScaDS.AI) Dresden/Leipzig · Leipzig, Germany +2

Reinforcement Learning Interpretability in RL

Nov 25, 2025

Reinforcement Learning to Initialize Newton-Raphson for AC Power Flow with Quantum Annealing-Based Environment Updates

The Newton-Raphson (NR) method is widely used for solving power flow (PF) equations due to its quadratic convergence. However, its performance deteriorates under poor initialization or extreme operating scenarios, e.g., high levels of renewable energy penetration. We propose the use of reinforcement learning (RL) to optimize the initialization of NR, and introduce a quantum-enhanced RL environment update mechanism that addresses the combinatorially large action space at each RL timestep by formulating the voltage adjustment task as a Quadratic Unconstrained Binary Optimization (QUBO) problem, solved with an Ising machine. RL initialization is benchmarked against flat start and start from the DC (linearized) PF solution on a standard 4-bus system, Iwamoto's ill-conditioned 11-bus system, and the IEEE 118-bus system under normal and stressed loading and reactive power limits, with verified operational solutions. On all systems, a supervised initializer refined by RL requires fewer NR iterations than flat and DC starts and than the same initializer without RL, for all seeds. For example, on the 118-bus system under normal and stressed loading, it reached 2.04 and 2.86 NR iterations, compared with 3.02 and 5.13 from DC start and 2.61 and 3.09 without RL. In wall-clock time, this pays off only for an initializer integrated into the solver and reused for many solves on a fixed topology. On the 4-bus system, a quantum-enhanced RL agent with a quantum-inspired annealer moved challenging initial states that required 29 and 44 NR iterations to initializations that required three NR iterations within one RL timestep.

Zeynab Kaseb, Matthias Moller, Lindsay Spoor +4

Electrical Sustainable Energy, Delft University of Technology, P.O. Box 5031, 2600 GA Delft, The Netherlands. · Applied Mathematics, Delft University of Technology, P.O. Box 5031, 2600 GA Delft, The Netherlands. · Leiden Institute of Advanced Computer Science, Leiden University, P.O. Box 9512, 2300 RA Leiden, The Netherlands. +3

Reinforcement Learning Quantum Annealing Learning to Optimize Power Systems

Nov 21, 2025

PepEVOLVE: Position-Aware Dynamic Peptide Optimization via Group-Relative Advantage

Macrocyclic peptides are an emerging modality that combines biologics-like affinity with small-molecule-like developability, yet their vast combinatorial space and multi-parameter objectives make lead optimization slow and challenging. Prior generative approaches for peptide lead optimization (e.g., PepINVENT) require chemists to pre-specify mutable positions for optimization, choices that are not always known a priori, and rely on static pretraining and optimization algorithms that limit the model's ability to generalize and effectively optimize peptide sequences. We introduce PepEVOLVE, a position-aware, dynamic framework that learns both where to edit and how to dynamically optimize peptides for multi-parameter improvement. PepEVOLVE (i) augments pretraining with dynamic masking and CHUCKLES shifting to improve generalization, (ii) uses a context-free multi-armed-bandit router that discovers high-reward residues, and (iii) couples a novel evolving optimization algorithm with group-relative advantage to stabilize reinforcement updates. During in silico evaluations, the router policy reliably learns and concentrates probability on chemically tractable sites that influence the desired peptide's properties. On a therapeutically motivated Rev-binding macrocycle benchmark, PepEVOLVE outperformed PepINVENT by reaching higher averaged permeability-weighted optimization scores of approximately 0.8 vs. 0.6, achieving best candidates with score of 0.97 (vs. 0.89), and converging in fewer steps under the task of optimizing permeability and lipophilicity with structure constraints. Overall, PepEVOLVE offers a practical, reproducible path to linear and macrocyclic peptide lead optimization when optimal edit sites are unknown, enabling more efficient chemical space exploration and improving design quality across multiple objectives to accelerate the design cycles.

Trieu Nguyen, Hao-Wei Pang, Shasha Feng

Merck & Co., Inc., Rahway, NJ, USA · University of South Florida, FL, USA

Molecular Optimization Drug Discovery Evolutionary Optimization Reinforcement Learning Multi-Objective Evolutionary Optimization

Nov 14, 2025

Humanoid Whole-Body Badminton via an Annealed Reinforcement Learning Curriculum

Humanoid robots have demonstrated strong capabilities for interacting with static scenes across locomotion and manipulation, yet dynamic real-world interactions remain challenging. As a step toward fast-moving object interactions, we present an RL training pipeline that yields a unified whole-body controller for humanoid badminton, coordinating footwork and striking without motion priors or expert demonstrations. In badminton, locomotion and striking are tightly coupled, making the final hitting objective difficult to optimize directly due to sparse rewards and conflicting gradients. We address this with an annealed curriculum that first stabilizes learning through auxiliary locomotion objectives, then progressively removes them to focus optimization on the final hitting objective. For deployment, we use an Extended Kalman Filter (EKF) to estimate and predict shuttlecock trajectories for target striking, and also develop a prediction-free variant that removes the EKF and explicit prediction. We validate the framework in simulation and on hardware. In simulation, two robots sustain a rally of 21 consecutive hits. In real-world tests with both machine-fed shuttles and human-robot rallies, the robot achieves outgoing shuttle speeds up to 19.1 m/s. Moreover, the prediction-free variant attains comparable performance to the EKF-based policy. Overall, our approach enables dynamic yet precise goal striking in humanoid badminton and suggests a path toward more dynamics-critical whole-body interaction tasks.

Chenhao Liu, Leyun Jiang, Ningyuan Tian +4

RL for Robotics Reinforcement Learning Humanoid Robot Control Curriculum RL Robotics

Nov 10, 2025

SpatialThinker: Reinforcing Scene Graph-Grounded Spatial Reasoning via Dense Rewards

Multimodal large language models (MLLMs) have achieved remarkable progress in vision-language tasks, but continue to struggle with spatial reasoning. Existing spatial MLLMs rely on large-scale datasets, explicit 3D inputs, architecture-specific modifications, or sparse Reinforcement Learning (RL) methods that provide insufficient guidance for spatially-grounded reasoning. We introduce SpatialThinker. To our knowledge, it is the first MLLM unifying Scene Graph Generation (SGG) and visual reasoning in a single pass via online RL. The model simulates human-like spatial perception by constructing a mental scene graph of task-relevant objects and relations, and reasoning toward an answer via dense spatial rewards. Our contributions are threefold: (1) SGG-grounded reasoning: integrating SGG directly within the reasoning chain rather than as a disjoint preprocessing step; (2) STVQA-7K: a high-quality spatial VQA training dataset via a scalable synthesis pipeline; and (3) a dense spatial reward design that enforces structured grounding during RL and generalizes to improve broad visual perception. SpatialThinker-7B achieves 3.6

\times

larger gains over SFT and

1.7\times

better in- and out-of-distribution generalization than sparse RL. Trained on only 7K samples, SpatialThinker-7B matches GPT-5 and outperforms GPT-4o, while SpatialThinker-30B surpasses both GPT-5 and Claude 4 Sonnet on average across 14 spatial and real-world benchmarks, demonstrating that structured spatial grounding with reward-aligned reasoning enables robust spatial understanding with limited data.

Hunar Batra, Haoqin Tu, Hardy Chen +3

University of Oxford · University of California, Santa Cruz

Multimodal Understanding Scene Graph Generation Visual Spatial Reasoning Reinforcement Learning RL for VLMs Visual Question Answering+5

Oct 31, 2025

Reinforcement Learning for Long-Horizon Unordered Tasks: From Boolean to Coupled Reward Machines

Reward machines (RMs) inform reinforcement learning agents about the reward structure of the environment, enabling support for non-Markovian tasks and improving sample efficiency. However, learning with RMs is ill-suited for long-horizon problems where subtasks can be completed in any order. In such cases, the amount of information to learn increases exponentially with the number of unordered subtasks. We address this issue by introducing three generalisations of RMs: (1) Numeric RMs allow users to express complex tasks in a compact form. (2) In agenda RMs, states are associated with an agenda that tracks the remaining subtasks to complete. (3) Coupled RMs have coupled states associated with each subtask in the agenda. In addition, we introduce QCoRM, a new task-decomposition Q-learning-based algorithm that leverages coupled RMs and preserves global optimality guarantees in tabular settings. Our experiments across four domains -- featuring both discrete and continuous action and state spaces -- demonstrate that QCoRM scales better than baseline algorithms for long-horizon problems with unordered subtasks.

Kristina Levina, Nikolaos Pappas, Athanasios Karapantelakis +2

Linköping University, Linköping, Sweden · Ericsson Research, Stockholm, Sweden

Reinforcement Learning Q-Learning Long-Horizon RL Reward Design for RL Non-Markovian RL

Oct 30, 2025

A Game-Theoretic Spatio-Temporal Reinforcement Learning Framework for Collaborative Public Resource Allocation

Public resource allocation involves distributing resources, including urban infrastructure, energy, and transportation, which are typically limited in capacity, to meet social demands. In real-world scenarios, resources are typically limited in capacity, which makes coordination among multiple resources essential. However, existing methods often optimize resource movements in an isolated manner and do not explicitly account for capacity-aware collaboration under spatio-temporal dynamics. To address this limitation, we introduce the Collaborative Public Resource Allocation (CPRA) problem, and propose a Game-Theoretic Spatio-Temporal Reinforcement Learning (GSTRL) framework to solve it. Our contributions are twofold: 1) We formulate CPRA as a potential game and construct the potential function based on the objective function of CPRA, laying a theoretical foundation for approximating the Nash equilibrium of this NP-hard problem; and 2) Our GSTRL framework effectively captures the spatio-temporal dynamics of the overall system. We evaluate GSTRL on two real-world datasets, where experiments show its superior performance. Our source codes are available at https://github.com/thunderlrr/GSTRL.

Songxin Lei, Qiongyan Wang, Yanchen Zhu +6

The Hong Kong University of Science and Technology (Guangzhou) · Peking University · Beijing Institute of Technology +1

Multi-Agent Coordination Reinforcement Learning Potential Games Cooperative MARL

Oct 28, 2025

Zero-Human Demonstration End-to-end Autonomous Driving with Trajectory Scorer

Human demonstrations are widely considered the cornerstone of end-to-end (E2E) autonomous driving despite human demonstration's scarcity for long-tail and safety-critical scenarios. Nonetheless, current E2E autonomous driving (AD) training paradigms continue to rely on human demonstrations. Imitation learning (IL) requires human demonstrations for training, whereas reinforcement learning (RL) has emerged as a promising alternative to reduce this dependency. However, most existing RL methods for E2E AD still rely implicitly on human demonstrations. A pure rewards-based RL method can overcome the need for human demonstrations, but general RL policy gradient methods suffer from the cold-start problem. In this paper, we propose ZTRS (Zero-human demonstration end-to-end autonomous driving with TRajectory Scorer) - a complete RL-based E2E planning paradigm trained solely on real-world images and rule-based rewards, entirely without human demonstration. Through our proposed Exhaustive Policy Optimization (EPO), a policy gradient variant tailored for enumerable trajectory actions and dense supervision, ZTRS enables the model to generalize better to long-tail driving scenarios. We demonstrate this generalization through our SOTA performance against IL approaches on both long-tail Navhard and closed-loop HUGSIM datasets. Project page: https://zhenxinli.net/ZTRS/.

Zhenxin Li, Nadine Chang, Wenhao Yao +9

Institute of Trustworthy Embodied AI, Fudan University · *Work done during an internship at NVIDIA. · NVIDIA +3

Long-Tail Autonomous Driving Reinforcement Learning End-to-End Autonomous Driving Autonomous Driving Autonomous Driving Planning RL for Autonomous Driving

Oct 27, 2025

Learning to Reason Efficiently with Discounted Reinforcement Learning

Large reasoning models (LRMs) often consume excessive tokens, inflating computational cost and latency. More broadly, in goal reaching sequential decision problems we often want to reach the goal quickly, and LRM reasoning can be viewed through this lens. We challenge the assumption that longer responses improve accuracy. By penalizing reasoning tokens using a discounted reinforcement learning setup (interpretable as a small token cost) and analyzing Blackwell optimality in restricted policy classes, we encourage concise yet accurate reasoning, analogous to preferring shorter successful trajectories in a stochastic shortest path problem. Experiments confirm our theoretical results that this approach shortens chains of thought while preserving accuracy.

Alex Ayoub, Kavosh Asadi, Dale Schuurmans +2

Amazon, University of Alberta · Amazon · University of Alberta

Reinforcement Learning RL for Language Model Reasoning Efficient Language Model Reasoning Policy Learning LLM Reasoning Large Reasoning Models

Oct 26, 2025

Toward Agents That Reason About Their Computation

While reinforcement learning agents can achieve superhuman performance in many complex tasks, they typically do not become more computationally efficient as they improve. In contrast, humans gradually require less cognitive effort as they become more proficient at a task. If agents could reason about their compute as they learn, could they similarly reduce their computation footprint? If they could, we could have more energy efficient agents or free up compute cycles for other processes like planning. In this paper, we experiment with showing agents the cost of their computation and giving them the ability to control when they use compute. We conduct our experiments on the Arcade Learning Environment, and our results demonstrate that with the same training compute budget, agents that reason about their compute perform better on 75% of games. Furthermore, these agents use three times less compute on average. We analyze individual games and show where agents gain these efficiencies. An interactive version of this paper is available at https://adrianorenstein.com/paper/toward-compute.

Adrian Orenstein, Jessica Jiajing Chen, Gwyneth Anne Delos Santos +2

Department of Computing Science, University of Alberta, Canada · Alberta Machine Intelligence Institute (Amii) · Department of Computer Science, University of Toronto, Canada +1

Reinforcement Learning Game-Playing Agents

Oct 22, 2025

On the Hardness of Reinforcement Learning with Transition Look-Ahead

We study reinforcement learning (RL) with transition look-ahead, where the agent may observe which states would be visited upon playing any sequence of

\ell

actions before deciding its course of action. While such predictive information can drastically improve the achievable performance, we show that using this information optimally comes at a potentially prohibitive computational cost. Specifically, we prove that optimal planning with one-step look-ahead (

\ell=1

) can be solved in polynomial time through a novel linear programming formulation. In contrast, for

\ell \geq 2

, the problem becomes NP-hard. Our results delineate a precise boundary between tractable and intractable cases for the problem of planning with transition look-ahead in reinforcement learning.

Corentin Pla, Hugo Richard, Marc Abeille +2

CREST, ENSAE · Criteo AI Lab · FairPlay Joint Team +1

Reinforcement Learning Model-Based Planning Markov Decision Processes

Oct 20, 2025

Consistent Zero-Shot Imitation with Contrastive Goal Inference

Zero-shot imitation learning requires an agent to reproduce expert behavior from a single demonstration without additional environment interaction or gradient updates at test time. We introduce Contrastive Inverse Reinforcement Learning (CIRL), a self-supervised framework for pre-training zero-shot imitation agents. Our methods rests on a key observation that many useful tasks can be summarized by a single goal state. We can thus convert the multi-task inverse RL problem into a more tractable goal-inference problem, and utilize state-of-the-art goal-conditioned RL methods to recover a policy that reaches the goal. During pre-training, CIRL jointly employs three components to learn without any rewards or demonstrations: (1) a variant of contrastive RL designed to learn maximum-entropy goal-conditioned policies, (2) an automatic goal proposal mechanism (GoalKDE) that drives exploration, and (3) a mean-field variational model that performs amortized goal inference from trajectories. We prove that this procedure consistently recovers the demonstrator's intent by accounting for the relative difficulty of reaching different states and show how structurally similar prior work may otherwise fail to infer the correct reward. Experiments on goal-conditioned and standard reward-maximizing control tasks show that CIRL outperforms prior zero-shot imitation methods, supporting the expressiveness of goals as a compact summary of behavior.

Kathryn Wantlin, Chongyi Zheng, Benjamin Eysenbach

Department of Computer Science, Princeton University, Princeton, NJ, USA.

Contrastive Learning Reinforcement Learning Imitation Learning Goal-Conditioned RL

Oct 14, 2025

Pretraining in Actor-Critic Reinforcement Learning for Locomotion

The pretraining-finetuning paradigm has facilitated numerous transformative advancements in artificial intelligence research in recent years. However, in the domain of reinforcement learning (RL) for robot locomotion, individual skills are often learned from scratch despite the high likelihood that some generalizable knowledge is shared across all task-specific policies belonging to the same robot embodiment. This work aims to define a paradigm for pretraining neural network models that encapsulate such knowledge and can subsequently serve as a basis for warm-starting the RL process in classic actor-critic algorithms, such as Proximal Policy Optimization (PPO). We begin with a task-agnostic exploration-based data collection algorithm to gather diverse, dynamic transition data, which is then used to train a Proprioceptive Inverse Dynamics Model (PIDM) through supervised learning. The pretrained weights are then loaded into both the actor and critic networks to warm-start the policy optimization of actual tasks. We systematically validated our proposed method with 9 distinct robot locomotion RL environments comprising 3 different robot embodiments, showing significant benefits of this initialization strategy. Our proposed approach on average improves sample efficiency by 36.2% and task performance by 4.3% compared to random initialization. We further present key ablation studies and empirical analyses that shed light on the mechanisms behind the effectiveness of this method.

Jiale Fan, Andrei Cramariuc, Tifanny Portela +1

Section of Microengineering, EPFL, 1015 Lausanne, Switzerland · Robotic Systems Lab, ETH Zurich, 8092 Zürich, Switzerland · ETH AI Center, ETH Zurich, 8092 Zürich, Switzerland

RL for Robotics Reinforcement Learning Robot Locomotion Actor-Critic Methods