Reinforcement Learning

Also known as RL

Follow this topic Ask about this topic

Momentum

144 papers in the last four weeks, up 243% on the four weeks before. 1.4% of all new papers.

Jul 13Week of Sep 28

Latest papers 1,138

Oct 7, 2026cs.LG

Temporally Interpretable Differentiable Decision Trees

Interpretability offers a solution to safe autonomy by providing transparency into an agent's underlying decision-making model. Within sequential-decision making tasks, differentiable decision trees (DDTs) are one approach to such interpretability, maintaining automatic-differentiable policies while providing humans with a discrete tree-based visualization. Nonetheless, current implementations of DDTs are not well-suited for sequential-decision making domains, as there exists an inherent mismatch between a tree's single-timestep behavior and a human's multi-timestep planning. Our work thus introduces time as a new dimension of interpretability, coined as temporal interpretability, and demonstrates how temporal abstractions via action chunking improve it. We achieve this by first introducing two novel policy gradient algorithms that incorporate action chunking. Additionally, to maintain parameter-efficient trees, we develop an information-theoretic tree restructuring algorithm that modifies the tree during training. Across four simulation environments, we find that warm-starting action chunked DDTs from a distilled action chunked policy is the most effective way to obtain temporally interpretable trees: they match neural network policies in three of the four domains while using up to 80

\%

fewer parameters. Our code is available at https://github.com/ei5uke/temp-interp.

Eisuke Hirota, Aarav Sane, Rohan Paleja

Department of Computer Science, Purdue University · Thomas Lord Department of Computer Science, University of Southern California

Interpretable ML Interpretability in RL Reinforcement Learning Policy Gradient Methods

Oct 7, 2026cs.LG

RollVerify: Bridging Efficiency and Accuracy in Long-Tail Rollout Reinforcement Learning

Reinforcement learning is crucial for improving large language models' reasoning and generalization. It relies on massive rollouts whose lengths become increasingly long-tailed as context windows grow. In on-policy training, these long-tail rollouts can result in GPU bubbles, reducing system utilization and limiting RL scalability. Asynchronous or partial-rollout methods improve throughput by relaxing synchronization, but inevitably introduce stale off-policy samples (trajectories) that may hurt final accuracy. Existing approaches mainly mitigate this off-policy issue by reweighting off-policy samples during training, yet they can still leave a performance gap compared to fully on-policy training. In this work, rather than passively reweighting samples during training, we propose RollVerify, a lightweight RL framework built on partial rollout that actively verifies and repairs samples before they enter training. Specifically, it introduces an off-policy shift metric OPS, to quantify the off-policy deviation of partially generated trajectories. Guided by the OPS constraint, RollVerify performs both sequence-level and token-level verification to identify and truncate invalid suffixes of trajectories. This yields high-quality samples that protect the models' accuracy while preserving the efficiency gains of partial rollout. Experiments on mathematical and tool-assisted mathematical reasoning show that RollVerify achieves accuracy comparable to on-policy training while reducing training cost. Additional code-generation results provide preliminary evidence beyond mathematics.

Yongqiang Yao, Jinru Tan, Kaihuan Liang +5

Shanghai Jiao Tong University · Central South University · SenseTime Research +2

RL for Language Models Reinforcement Learning RL for Language Model Reasoning

Oct 7, 2026math.OC

Boundary-aware Reinforcement Learning for Hypercube State Spaces via Deterministic Policy Gradient

We develop a continuous-time deterministic policy gradient framework for reinforcement learning with reflected state dynamics, where the state process is governed by a controlled reflected stochastic differential equation on a hypercube. Under suitable regularity assumptions, we establish the connection between the value function and the Neumann Bellman equation, introduce an advantage-rate function that yields a deterministic policy gradient formula, and prove the martingale characterization theorem. Motivated by these theoretical results, we propose a continuous-time deep deterministic policy gradient algorithm for reflected stochastic systems, in which the Neumann boundary condition is imposed via either soft penalization or hard architectural constraint. We further quantify the discrepancy between the ideal continuous-time dynamics and the discretely sampled exploratory dynamics executed in practice, showing that the error decays as the time grid is refined and exploration noise vanishes. Our experiments on reservoir control problems illustrate the effectiveness of the RL framework, highlighting that boundary-aware methods substantially reduce Neumann boundary residuals and enhance learning stability.

Lijun Bo, Yijie Huang, Chenhao Lu

RL Control Reinforcement Learning Continuous-Time RL Deep Deterministic Policy Gradient Stochastic Differential Equations

Oct 7, 2026cs.LG

CERO: Where and When to Allocate Rollouts for RL Post-Training

Adaptive rollout methods for group-relative reinforcement learning typically allocate a fixed per-update budget across prompts. We instead study how to coordinate a finite rollout budget over the entire training horizon. We formulate this problem using a concave surrogate utility of cumulative prompt exposure and introduce CERO, an online primal dual scheduler for prompt admission and budget pacing. In our experiments, each admitted prompt receives a fixed-size response group. CERO instead adapts which prompts are selected, how often they are revisited across rounds, and how many groups are generated in each round. A compact Fenchel representation linearizes the dependence on cumulative exposure, while projected online gradient descent updates prompt-specific supporting slopes and a shared budget price using reward-variation feedback and budget deviations. We establish pathwise guarantees for the surrogate allocation objective against fixed-rate and same-path time-varying benchmarks, with explicit terms for proxy discrepancy and rate variation. Under matched training-response budgets, CERO attains the highest avg@16 macro-average on each of three backbones across five mathematical reasoning benchmarks. Mechanistic analyses link CERO's prompt choices to within-group reward contrast, while multi-seed ablations show gains from adaptive pacing over both uniform and preset spending schedules.

Yiming Zong, Yige Wang, Xing Hu +2

Department of Industrial Engineering & Decision Analytics, Hong Kong University of Science and Technology · Innovation and Information Management, The University of Hong Kong · Faculty of Engineering & Faculty of Business and Economics, The University of Hong Kong

Reinforcement Learning RL Post-Training RL for Language Model Reasoning Group Relative Policy Optimization Primal-Dual Optimization

Oct 6, 2026cs.AI

AGAR: a reinforcement learning substrate for LLM program evolution

Given a task and an evaluator, a language model can rewrite a candidate program while a search loop decides which rewrites survive, offering a practical route to algorithm discovery. But that loop is governed by five constants set by hand: which parent to select, how hard to mutate, how to keep diversity, what to remember, and a scalar score that never says which part of the program earned it. Reinforcement learning already has an estimator for each. The obstacle is that program evolution is not usually written down as a decision process. We formalize it as a Markov decision process whose action is the modular prefix the model is conditioned on, rather than the program it emits. Credit assignment, value estimation, adaptive exploration, and experience memory can then attach to distinct components. AGAR (Algorithm Generation As RL) provides the resulting substrate: any estimator can be replaced or switched off without changing the controller, making the transfer auditable one mechanism at a time, with no gradient training of the backend model. Across 19 tasks, two backends, and three seeds under one harness, AGAR improves on the stronger of two published baselines on most tasks, with gains concentrated in the competitive-programming family. The formalization also yields a checkable reading of prior work: these systems are implicitly zero-discount, not by choice, but because fitness is exogenous to an individual rather than a return over successors, leaving a discount factor nothing to act on.

Haoran Li, Zengle Ge, Xiaomin Yuan +11

University of the Chinese Academy of Sciences · Baidu · Tsinghua University +2

LLM-Based Program Synthesis RL for Code Generation Reinforcement Learning Credit Assignment in RL

Oct 6, 2026cs.LG

Convex-Concave Reinforcement Learning

Policy learning drives many of the most consequential and heavily-invested applications of reinforcement learning today. Yet the core optimization problem it rests on (maximizing expected return) is notoriously non-convex, even under a direct policy parameterization, and the field has largely responded by avoiding it: optimizing convex surrogate approximations of the return under trust-region constraints (NPG, TRPO, PPO, AWR). We show that this seemingly unstructured problem is not actually structureless. In log-density-ratio coordinates

y := \log[π/π_n]

, the exact per-iteration objective, computable via per-decision importance sampling (PDIS), is a difference-of-convex-constrained difference-of-convex (DC-constrained DC) program. This structure lets us move beyond surrogate approximations: it recovers CPI, NPG, TRPO, and AWR as special cases along interpretable axes, and it opens a multi-step axis

k

that couples consecutive decisions. We solve the per-iteration program with sequential convex programming (SCP), the standard solver for difference-of-convex problems, and give convergence guarantees under mild conditions, bridging the difference-of-convex optimization and RL literatures. Empirically, multi-step Convex-Concave RL (CCRL) wins on diagnostic MDPs where credit must propagate across a horizon (its advantage growing with the dependency length), is competitive with a tuned PPO on classic control, and on a realistic, stochastic, mid-horizon healthcare domain converges markedly faster than tuned PPO to the same near-optimal survival, with an 11.3% higher area under the training curve.

Shripad V. Deshmukh, Yaswanth Chittepu, Dhawal Gupta +2

University of Massachusetts · Institute of Foundation Models

Reinforcement Learning Policy Gradient Methods Nonconvex Optimization Step-Level Credit Assignment in RL

Oct 6, 2026cs.LG

LASER: Latent Space Adjoint Matching for Support-Constrained Entropy-Regularized Offline RL

While offline reinforcement learning (RL) enables policy optimization from static datasets without costly online interaction, it remains bottlenecked by the risk of executing out-of-distribution (OOD) actions. Recent approaches mitigate this by learning a behavior-cloning policy through flow matching and then performing RL within its constrained latent space. However, naively optimizing the latent policy can easily cause the policy to collapse into a brittle mode or exploit sharp artifacts of the learned critic. In this work, we find that entropy regularization is essential in latent-space RL for addressing these challenges. We introduce LASER, a novel offline RL algorithm that applies latent-space adjoint matching to achieve entropy-regularized latent-space RL with expressive flow policies while avoiding backpropagation through time. Through comprehensive experiments on 40 challenging OGBench tasks with varying dataset qualities, we show that LASER achieves state-of-the-art performance. Notably, LASER uses fixed method-specific hyperparameters across all tasks and outperforms the evaluated baselines, including those with task- and dataset-specific tuning, which highlights the robust applicability of LASER. Project website: https://mit-realm.github.io/laser/.

Songyuan Zhang, Oswin So, Eric Yang Yu +3

MIT · MIT Lincoln Laboratory

Reinforcement Learning Offline RL Maximum Entropy RL

Oct 6, 2026cs.LG

Reinforcement Learning with Conformal Action Sets: An Application to Sequential Recommendation

Sequential recommenders typically use a fixed slate size even though the number of useful alternatives changes within a session. We propose Reinforcement Learning with Calibrated Pruning (RLCP), which adapts the retained action set using critic scores and an online threshold. The threshold is updated from binary feedback indicating whether the set contains an action in a proxy target. We prove a deterministic bound on the observed proxy miss rate along adaptive trajectories. To quantify the effect of pruning on reward, we derive an exact decomposition of value loss into filtering and selection losses. Under explicit proxy and critic approximation conditions, this decomposition yields a finite session reward bound that also accounts for imperfect selection and set truncation, without requiring the learning parameters to converge. Experiments on KuaiRand-Pure and MovieLens 1M compare two RLCP implementations with four RL baselines. In each of the 19 configurations, at least one RLCP variant achieves the highest catalog diversity, reaching

1.11\times

to

5.21\times

that of the strongest baseline, with competitive session depth and no larger retained sets.

Wenwen Si, Honghao Wei

University of Pennsylvania · Washington State University

Sequential Recommendation Reinforcement Learning Conformal Prediction Recommender Systems

Oct 6, 2026cs.CV

Selective Transfer of RL Updates for Visual Reasoning

Model merging provides a training-free way to transfer reasoning capabilities from language models to vision-language models (VLMs), but endpoint-based transfer can conflate pre-existing model differences with changes acquired during reasoning post-training. We instead formulate capability transfer around the training-stage update, isolating the parameter changes induced by reinforcement learning (RL). Yet transferring this update in full remains suboptimal: we find that its components differ substantially in cross-model transferability, with dominant directions transferring more effectively than the complete update. Based on this finding, we introduce Selective-RL, which isolates the RL-stage update, retains its dominant matrix-wise directions with magnitude preservation, and transfers them to the language modules of a VLM. Across three model families and five visual-reasoning benchmarks, Selective-RL improves full-update interpolation in 12 of 15 comparisons, including an 8.55 percentage-point MathVision gain on the Qwen recipient. Matched controls show that update magnitude or arbitrary low rank alone does not reproduce these gains. These results highlight a distinction between what is acquired during post-training and what remains transferable across models, providing a training-stage perspective on cross-model capability transfer. Code is available at https://anonymous.4open.science/r/selective-rl.

Suxin Ji, Hungtao Wan, Mingjun Liu +1

University of Pennsylvania · Independent Researcher · University of Electronic Science and Technology of China +1

Reinforcement Learning VLM Reasoning RL for VLMs

Oct 6, 2026cs.LG

Reinforcement Learning with Segment Reward Feedback under Linear Function Approximation

Classical reinforcement learning (RL) assumes that a reward is observed for every visited state-action pair. However, in real-world applications such as autonomous driving, such fine-grained feedback can be costly or difficult to collect, whereas trajectory-level feedback may be too sparse for efficient learning. To provide a general feedback model bridging these two extremes and handle large state spaces, we study RL with segment reward feedback under linear function approximation. Our work answers how the granularity of segment feedback and the choice of segmentation influence learning. For equal-length segments with known transitions, we design algorithms

\bitssegd

and

\edlinucbsegd

for binary and sum feedback types, respectively. They adopt posterior sampling with planning to achieve computational efficiency and the E-optimal experimental design to attain near-optimality. Nearly matching lower bounds are established. For equal-length segments with unknown transitions, we develop a unified

\seglsvits

framework with two instantiations for binary and sum feedback, which carefully integrates the posterior estimated reward parameters into least-squares value iteration. These results reveal a fundamental insight: under binary feedback, increasing the number of segments significantly reduces the regret through an exponential factor, while surprisingly, under sum feedback, the granularity of segments does not affect learning much. Finally, to investigate whether segmenting according to state-action features can further expedite learning, we design an algorithm

\uneqsegbitsd

that allows arbitrary segmentations. The resulting regret bound shows that under the usual elliptical potential analysis, the influence of state-action features on the regret appears only through logarithmic factors, and equal segmentation achieves the best performance.

Fengxu Liu, Siwei Wang, Gal Dalal +2

National University of Singapore. · Microsoft Research Asia. · NVIDIA Research. +2

Regret Minimization in RL Reinforcement Learning

Oct 6, 2026cs.LG

Revisiting Temporal Regularization for Smooth Control in Deep Reinforcement Learning

Deep Reinforcement Learning policies can produce nonsmooth action oscillations that hinder deployment on physical robots. Existing architectural and penalty-based approaches seek spatial smoothness by directly reducing sensitivity to changes in state inputs, but their broad constraints can degrade task performance as stronger smoothing is pursued. Temporal regularization instead constrains action differences along observed transitions, but has been considered unable to provide the spatial smoothness needed under observation noise. We revisit this assumption by proving that the temporal penalty bounds the expected action differences between current states sharing a next state, revealing a spatial effect that empirically extends to spatial smoothness. Building on this finding, we propose Conditioning for Action using only Temporal Smoothness (CATS), which combines a temporal penalty with linear ramp-up. We highlight temporal regularization's ability to provide spatial smoothness while better preserving task performance than explicit spatial regularization. Through linear ramp-up, CATS allows the policy to learn rewarding behavior before progressively smoothing its actions, improving return preservation and both temporal and spatial smoothness. Experiments in both simulation and the real world show that CATS substantially reduces action oscillation without degrading task performance, with little computational overhead.

SungJae Ahn, Jeong Woon Lee, Kyoleen Kwak +1

College of Software, Kyung Hee University.

RL Control Reinforcement Learning Robotic RL

Oct 6, 2026cs.LG

FC-SWE: Failure-Conditioned RL for Long-Horizon Software Engineering Agents

Repository-level software engineering (SWE) is a challenging long-horizon setting: agents must reason over extended interactions, use tools, and adapt to stateful environments. Recent work trains SWE agents with reinforcement learning methods such as Group Relative Policy Optimization (GRPO), which independently sample multiple trajectories per issue, test the resulting patches, and compare terminal rewards within a fixed group. However, this training setup does not reuse verifier feedback from failed patches as context for subsequent attempts, even though this feedback contains valuable diagnostic information about what went wrong. Training on recovery trajectories is challenging because the preceding outcome determines whether the next trajectory is generated, while the failed execution determines its conditioning context. We introduce FC-SWE, a failure-conditioned RL framework that incorporates recovery attempts into policy training. After a patch fails verification, FC-SWE restores the repository to its original task state and uses the failed patch and verifier feedback as context for a recovery trajectory. FC-SWE adapts GRPO to these chains of complete, multi-turn tool-use trajectories through two mechanisms. Trajectory-local rewards preserve each attempt's verifier outcome, preventing recovery success from rewarding an earlier failed patch. Active-set advantage estimation forms a comparison group from all initial and recovery trajectories actually executed for the same issue, so failed attempts remain in the group while unexecuted attempts are excluded. On all 500 SWE-bench Verified tasks under a verifier-assisted protocol, FC-SWE with Qwen3.5-4B and SWE-agent achieves 41.7% Resolved@1 and 52.8% Resolved@2, compared with 38.9% and 48.5% for GRPO. Although trained with at most two attempts per chain, FC-SWE reaches 70.7% Resolved@11 under an eleven-attempt test-time budget.

Jia Liufu, Bin Hu, Linglin Jing +5

NVIDIA · Tsinghua University

RL for Code Generation Software Engineering Agents Reinforcement Learning Reinforcement Learning with Verifiable Rewards RL for Tool Use

Oct 6, 2026math.NA

Learned Adaptive Multiresolution Diffusion Imaging

Adaptive multiresolution methods reduce representation cost by concentrating fine-scale degrees of freedom where needed, but their tree updates are usually governed by fixed local criteria. We introduce Learned Adaptive Multiresolution Diffusion Imaging (Learned AMDI), which preserves the AMDI fixed-tree propagator and hierarchy constraints while replacing the post-propagation selector with a shared local policy trained by proximal policy optimization. Regression tests reproduce deterministic AMDI trajectories to machine precision when identical trees are used. In the Haar implementation studied here, the deterministic one-step selector accepts no refinements in 54 decisions. Across nine held-out cases, Learned AMDI executes 393 refinements and reduces the mean terminal reference discrepancy from

0.17496

to

0.13657

, while occupancy rises from

0.13737

to

0.26660

. Step-resolved diagnostics reveal occasional small adaptation-energy increases; fixed-tree energy stability therefore does not guarantee monotonicity of the learned outer iteration. At comparable occupancy, a validation-tuned observed-detail threshold reaches a discrepancy of

0.13792

with slightly better RMSE and SSIM, placing both methods on essentially the same accuracy--occupancy tradeoff. A decision-1-only control reaches

0.13742

, indicating that most of the improvement on this static benchmark arises from the initial allocation. The shared actor transfers without retraining to

64\times64

and

128\times128

images, improving reference discrepancy, RMSE, and SSIM relative to deterministic AMDI, while the frozen threshold rule remains competitive. Learned AMDI thus provides a hierarchy-constrained, resolution-transferable mechanism for adaptive allocation and clarifies the contribution of sequential decisions.

Christian Tantardini, Stig Rune Jensen, Roberto Di Remigio Eikås +1

Center for Integrative Petroleum Research, King Fahd University of Petroleum and Minerals, Dhahran 31261, Kingdom of Saudi Arabia · Hylleraas Centre for Quantum Molecular Sciences, Department of Chemistry, UiT The Arctic University of Norway, N-9037 Tromsø, Norway · Algorithmiq S.r.l., c/o STLex, Via della Chiusa 15, 20123, Milano, Italia

Reinforcement Learning Adaptive Mesh Refinement

Oct 5, 2026cs.LG

Adapting to Changes in Agent Behavior via Finite-Depth Policy Sensitivity

Adapting a reinforcement learning policy to changes in another agent's behavior typically requires a large amount of new interaction data. Policy sensitivity provides a first-order prediction of how a locally optimal policy changes with a behavioral parameter, but its computation requires second-order derivatives whose effects propagate across future interactions. We develop a finite-depth framework to estimate this sensitivity by approximating the policy Hessian and mixed derivative using information from a reference environment. The method features an adjustable propagation depth which determines where derivative propagation along the trajectory is truncated. We characterize the derivative contributions omitted by finite-depth propagation and derive truncation-error bounds for the approximated derivatives and resulting policy sensitivity. The bounds are nonincreasing with propagation depth and vanish at full-horizon propagation. Using a belief-driven pursuit-evasion game as a validation scenario, the proposed method generally achieves lower derivative-estimation errors as the propagation depth increases and outperforms the baseline methods in both estimation accuracy and policy adaptation. The sensitivity-based initialization improves zero-shot return over direct transfer, and also shows advantages for the subsequent fine-tuning in the target environment.

Lan Shi, Daigo Shishika, Xuan Wang

George Mason University, USA

Reinforcement Learning Multi-Agent Reinforcement Learning

Oct 5, 2026cs.LG

Considering Context: When World Models Need Context Encoders

Methods for generalization in model-based reinforcement learning typically assume that an agent cannot recover the latent context governing the environment dynamics from its own experience, and therefore supplies it externally. We formalize and test this assumption with \emph{predictive sufficiency}, which quantifies what access to the context adds to next-step prediction under the visitation distribution an agent induces, and separates that quantity into a history-recoverable part, a residual requiring the true context, and the deficit added by a finite model. We classify context-aware algorithms by the predictive risk their conditioning set can target and demonstrate across environments of increasing identification difficulty that the headroom does not follow the MDP class. The same task under different priors leaves predictive headroom in one setting and nothing distinguishable from zero in another, where the agent's behavior implicitly identifies the context and any benefit of such a mechanism cannot be attributed to missing information. Where headroom persists, the learned state exposes it only partially, and adding the true context still lowers the risk. Our contribution is a practical criterion for matching contextual mechanisms to the information available to them, estimated from the ordinary trained agent without a reference policy.

Oleg Smirnov, Sofiane Ennadir, John Pertoft +2

King AI Labs, Microsoft Gaming

Reinforcement Learning World Model Learning Generalization in RL Model-Based RL

Oct 5, 2026cs.AI

GPlaceRL: An Open-Source Graph Reinforcement Learning Framework for Detailed Placement

Reinforcement learning (RL) has emerged as a promising approach for placement optimization, particularly when combined with graph neural networks (GNNs) that capture circuit connectivity. However, most learning-based placement approaches focus on floorplanning, macro placement, or global placement, while detailed placement refinement remains relatively unexplored. In this paper, we present GPlaceRL, an open-source graph reinforcement learning framework for detailed placement refinement. GPlaceRL represents legalized placements as graphs and provides a modular environment for studying graph encoders, policy architectures, reward formulations, and local placement actions. To demonstrate the capabilities of GPlaceRL, we conduct a systematic evaluation of proximal policy optimization (PPO) policies with graph attention network (GAT) encoders in a per-design optimization setting. Across five placement benchmarks, the best greedy evaluation results achieve HPWL improvements ranging from

3.27\%

to

32.87\%

. The results highlight the importance of compact GAT architectures and flexible local action spaces for placement optimization. Overall, GPlaceRL provides a reproducible and extensible framework for systematic research on RL-based detailed placement refinement.

Pavlos Stoikos, Foteini Oikonomou, Christos Poulos +5

Department of Electrical and Computer Engineering, University of Thessaly, Volos, Greece · School of Computing Science, University of Glasgow, UK · Department of Electronic and Electrical Engineering, Trinity College Dublin, Ireland

Reinforcement Learning Electronic Design Automation RL for Combinatorial Optimization

Oct 5, 2026cs.RO

I-BFM: Reward-Conditioned Robust Humanoid Interaction via Unsupervised Reinforcement Learning

Behavioral foundation models (BFMs) have recently shown that a single humanoid policy can support diverse whole-body control, but extending such generality to physical interaction remains challenging. We introduce I-BFM, to our knowledge the first BFM for humanoid-object interaction. Rather than relying on task-specific policies or reference tracking, I-BFM learns a shared representation of the coupled dynamics among the humanoid, objects, and their contacts using forward-backward representations and unsupervised reinforcement learning. Given a downstream task reward, the same policy can be directly conditioned on a latent command to execute closed-loop interaction without task-specific policy optimization. To improve interaction control over different time scales, we further train the policy with both short-horizon interaction targets and longer-horizon goal targets. A single I-BFM policy performs carrying, pushing, and kicking, while also supporting goal reaching, motion tracking, stylistic control, and long-horizon task chaining. More importantly, it remains effective after large deviations from nominal execution: on Carry, I-BFM achieves 94.3% nominal success and retains 89.3% success after robot falls, compared with 1.3% for a planning-based baseline. Real-world experiments on a Unitree G1 further demonstrate diverse loco-manipulation behaviors, rapid recovery from interaction failures and external disturbances, and task chaining without task-specific retraining.

Ziqi Han, Yitang Li, Junhan Sun +8

Tongji University · Tsinghua University · Zhejiang University +3

Representation Learning Reinforcement Learning Humanoid Robot Control Robust RL Human-Object Interaction

Oct 5, 2026cs.LG

Fast Last-Iterate Convergence in Zero-Sum Markov Games with Bandit Feedback

We study last-iterate convergence in unknown two-player zero-sum discounted Markov games with bandit feedback. The players learn independently along a single trajectory without observing each other's actions. We develop Adaptive Regularized TD Learning (ARTD), which achieves a

\widetilde{\mathcal{O}}(t^{-1/4})

duality gap bound for the current policies under a uniform hitting time assumption, with high probability simultaneously over all rounds and starting states. This improves the

\widetilde{\mathcal{O}}(t^{-1/(9+ν)})

rate of Cai et al. (2023), for any fixed

ν>0

, under the same feedback model and hitting time assumption. Our algorithm requires no knowledge of the hitting time bound, the time horizon, or the confidence level. To stabilize policy learning as value estimates change, we separate fast temporal difference averaging from bounded value updates. We adapt log-barrier regularization to the progress of value estimation, controlling both policy and value errors throughout learning. Together, these mechanisms enable fast convergence of the policies actually played, even when the players learn independently from bandit feedback.

Yuheng Zhang

University of Illinois Urbana-Champaign

Last-Iterate Convergence Reinforcement Learning Zero-Sum Games Temporal-Difference Learning

Oct 5, 2026cs.CL

HuatuoGPT-3: RL-Only Domain Adaptation from Base Models

Domain adaptation aims to turn a general-purpose large language model (LLM) into an expert for a target domain. While the dominant SFT+RL pipeline offers a convenient cold start, it may reduce exploration diversity and introduces additional complexity through multi-stage optimization. These limitations motivate RL-only adaptation. However, pure on-policy RL suffers from a cold-start problem, while mixed-policy RL still falls short: informative tokens in teacher outputs are learned too slowly in early training, and stale teacher outputs can hinder later improvement. We identify these two failure modes as Gradient Starvation and Teacher-Distribution Anchoring. To address them, we propose One-stage Policy Optimization (OnePO), which treats teacher outputs as transient guidance for policy improvement. OnePO combines Adaptive Objective Evolution to strengthen learning on informative low-probability teacher tokens and Teacher Retirement to discard teacher outputs once the current policy can surpass them. On medical adaptation, OnePO achieves 67.2 on HealthBench (Total) with only 20K training samples, outperforming SFT+RL and pure RL by 2.7 and 7.4 points, respectively. We further scale OnePO to produce HuatuoGPT-3, an open-source medical LLM series whose 27B variant reaches 70.1 on HealthBench (Total) and 71.4 on HealthBench Professional, surpassing frontier models such as GPT-6 Astra. Models and code are available at https://github.com/FreedomIntelligence/HuatuoGPT-3.

Junying Chen, Xinyuan Xie, Ziniu Li +7

The Chinese University of Hong Kong, Shenzhen · Shenzhen Research Institute of Big Data · Shenzhen Loop Area Institute +1

RL for Language Models Reinforcement Learning Domain Adaptation for LLMs Medical LLMs

Oct 4, 2026cs.LG

An LLM-in-the-loop RL Framework for Bioinformatics Feature Selection

High-dimensional bioinformatics data, characterized by a large number of features relative to the number of samples, pose major challenges such as the ``curse of dimensionality,'' leading to overfitting, high computational cost, and poor generalization. Traditional feature selection methods often suffer from limited scalability and adaptability in such domains. We propose an LLM-in-the-loop reinforcement learning (RL) framework for bioinformatics feature selection, where the RL agent formulates feature selection as a sequential decision-making task, while the large language model (LLM) enhances the process in two ways: (1) guiding exploration through domain-informed advice, and (2) providing hybrid rewards that integrate data-driven performance with knowledge-driven evaluation. The LLM also produces explanations to improve interpretability for human experts without altering the RL policy update. Experiments on diverse bioinformatics datasets show that the LLM-in-the-loop framework outperforms baselines, achieves stable performance across downstream models, and converges faster than pure RL.

Xinyuan Wang, Deepti Agrawal, Yanjie Fu

School of Computing and AI Arizona State University Tempe, USA

Reinforcement Learning LLM-Guided RL

Oct 4, 2026cs.AI

Factoriax: A GPU-Accelerated Factorio-Style Simulator for Reinforcement Learning

We introduce Factoriax, a GPU-accelerated factory-building simulator written in JAX. In Factoriax, an agent must collect resources, build machines using those resources, and then automate the collection and crafting process by arranging machines on the map to build production pipelines. This paper discusses the structure of the Factoriax simulator and an initial benchmark called Easy Rocket in which an agent is tasked with building a resource-intensive machine called the Rocket in a limited number of game ticks to escape the planet. We also publish results from a number of PPO-based training runs on Easy Rocket. Factoriax is built to be fast. A 1-billion-step PPO training run, equivalent to 500,000 episodes, runs on Easy Rocket in about 8 minutes on a single NVIDIA A100. A standard laptop GPU can complete the same run in about 84 minutes. Our trained PPO agent learns to gather resources, craft machines from those resources, and place them on the map through a curriculum reward directly tied to a manually designed set of achievements. After training, the agent does not place machines in a functional spatial configuration, failing to fully complete the benchmark, and leaving the challenge open for future attempts.

Mickey Beurskens, Tristan Tomilin, Thiago D. Simão

Eindhoven University of Technology

RL Benchmarks Reinforcement Learning

Oct 4, 2026cs.AI

Hierarchical Reinforcement Learning with Stable Temporal Abstraction for Language Model Agents

Hierarchical reinforcement learning improves long-horizon control by organizing primitive actions around persistent subgoals and assigning credit at multiple temporal scales. Recent hierarchical language agents bring these benefits to interactive tasks by explicitly separating subgoal planning from action execution. We observe, however, that an explicit hierarchy does not by itself determine how stable the resulting temporal abstraction is: the learned boundary policy may replace the subgoal almost every turn, making it effectively transient, or retain a subgoal after it has stopped being appropriate. We call this temporal abstraction instability. We propose Stable Temporal Abstraction via Constrained Optimization (STAC), a constrained boundary-policy optimization method that represents premature replanning and stale persistence as constraint costs. STAC applies the resulting Lagrangian costs only to the sampled boundary decision, leaving the underlying algorithm's rewards, critic targets, subgoal advantages, and primitive-action advantages unchanged. Across two backbones and two benchmarks, STAC improves success over a strong hierarchical baseline by

8.1

and

7.9

points on ALFWorld and WebShop with Qwen3-0.6B, and by

23.5

and

15.8

points with Llama-3.2-1B-Instruct.

Shayan Mohajer Hamidi, Yize Cheng, Yuanda Xu +2

LinkedIn Corporation · University of Maryland

Reinforcement Learning Hierarchical RL RL for LLM Agents

Oct 4, 2026cs.LG

Task Inference Beyond Least Squares in Behavioral Foundation Models

Behavioral Foundation Models (BFMs) aim to solve a wide range of downstream tasks without test-time policy learning by inferring a task vector from the reward function. While efficient, the retrieved policies are often suboptimal because of how this task vector is inferred, typically with ordinary least squares (OLS). OLS minimizes reward reconstruction error but leaves the ordering of rewards unconstrained, which can bias the successor measure of the retrieved zero-shot policy away from that of the optimal policy. In this work, we propose BLS, an efficient test-time inference method that balances minimizing reward reconstruction error with reducing successor-measure mismatch. Theoretically, we provide a suboptimality gap upper bound characterized by both successor-measure and reward-function residuals. Empirically, we evaluate BLS on top of state-of-the-art BFMs across benchmarks for locomotion, manipulation, and humanoid control. BLS outperforms existing task inference baselines with negligible computational overhead. Project page: https://embodiedai-ntu.github.io/BLS

Kuan-Hsun Tu, Chien-Sheng Chiang, Hsin-Wei Chen +2

National Taiwan University

Reinforcement Learning Multi-Task RL

Oct 4, 2026cs.LG

On Semi-Markov Suboptimality in Hierarchical Reinforcement Learning

Hierarchical reinforcement learning uses temporally extended subtasks for exploration, yet committing to their execution can restrict both deployment and policy learning. We identify and separate the resulting execution and policy suboptimality. Task and execution trees distinguish reward objectives from policy choices and decision interruption. A Unified Value Function for HRL and a four-stage Generalized Hierarchical Bellman Equation then support a common analysis of both losses. Under bounded rewards and uniform termination, we establish hierarchical policy and execution improvement results. With the remaining node policies fixed, task-subtree compatibility and node-policy optimality under the original execution mode establish when Markov execution is optimal. The resulting decomposition leads to independent execution choices for behavior, targets, and deployment. We instantiate this principle through execution improvement and one-stage or two-stage policy improvement at arbitrary hierarchy depth. Option-based and goal-conditioned experiments demonstrate complementary gains from changing execution and changing the learning target. Controlled stochastic environments show how these gains depend on stochastic transition strength and spatial structure. This framework makes execution design an explicit component of hierarchical policy optimization.

Bingyun Liu, Yuheng Jing

Institute of Automation Chinese Academy of Sciences

Reinforcement Learning Sequential Decision Making Hierarchical RL

Oct 4, 2026cs.CV

Long-MDR: Long-Context Reinforcement Learning for Multimodal Deep-Research Agents

The next generation of multimodal research agents must reason over long-lived research histories rather than short model completions. During a single task, an agent may repeatedly search the web, inspect visual evidence, revisit earlier hypotheses, and accumulate tens of thousands of tokens of multimodal context. Despite this trend, online RL for multimodal research agents remains largely confined to shorter contexts and interaction horizons. We push online RL training to 128k context and 75+ tool-interaction turns. To our knowledge, this is the first online multimodal deep-research RL study trained at 128k context, and the first trained with a 75 tool-turn horizon. Scaling to this regime exposes several practical limitations of conventional RL training. Early in training, weak policies make poor use of large interaction budgets, causing expensive rollouts with little reward improvement. Later, policy entropy can collapse before performance has saturated, prematurely ending useful learning. We introduce Long-MDR, a three-component training recipe designed specifically for this setting: On-Policy Distillation Warmup, Progressive Horizon Expansion, and Entropy-Triggered Rescue. Together, these techniques improve both the learning efficiency and stability of long-horizon RL, enabling continued gains in a regime where direct training is slow and costly. At a 50-turn evaluation budget, our RL-trained Long-MDR-9B ranks first on five of six benchmarks among the compared 7B-9B agents.

On Tai Tang

The Hong Kong University of Science and Technology (HKUST)

Reinforcement Learning Deep Research Agents Multimodal Agents RL for LLM Agents

Oct 4, 2026cs.AI

Beyond Instruction Following: Learning Grounded Skill-Following with Skill Contracts

Instruction following typically enforces discrete, response-level requirements, whereas an expert-authored skill prescribes procedural requirements spanning multiple phases and environment interactions. Given such a skill, we train the executor to execute all required phases instead of focusing solely on the final answer. We therefore introduce Grounded Skill-Following, which requires an agent to execute a fixed, expert-authored skill across its required phases by grounding decisions in environment observations. To achieve verifiable procedural execution, we formulate each skill as a skill contract combining visible skill instructions with an explicit contract runtime. The runtime specifies required phases, admissible actions, permitted transitions, and accepted termination. This structure provides a dense, verifiable training signal throughout execution. We leverage this by introducing Verified Progress Credit, which assigns rewards upon the initial completion of contract milestones and aggregates them into the trajectory return to guide policy optimization. During rollout, the contract runtime continuously tracks state transitions to provide Contract-State Feedback, which indicates whether the latest action is accepted and guides the agent toward valid next actions. To measure procedural compliance, we introduce the Protocol Completion Rate (PCR), defined as reaching accepted termination through all required phases, and decouple it from the final Task Outcome. Jointly trained with our framework, Qwen3.5-4B achieves Protocol Completion Rates of 99.27% on Math and 99.96% on Search, while slightly outperforming original baselines in Task Outcome (82.95% and 46.61%, respectively). Controlled studies examine how skill instructions, training signals, and contract-state feedback affect both metrics, while withholding interventions evaluate behavioral dependence on observation content.

Jianghan Shen, Zhenjie Liu, Yue Li +10

Nanjing University · Shanghai Artificial Intelligence Laboratory · Peking University +3

Reinforcement Learning LLM Agent Skill Learning Reinforcement Learning with Verifiable Rewards RL for LLM Agents

Oct 4, 2026cs.LG

Direction-Conditioned Policies for Online Goal-Conditioned Reinforcement Learning

Contrastive Reinforcement Learning (CRL) learns representations that estimate goal reachability, yet its policy remains conditioned on raw goals and therefore does not directly exploit the geometry encoded by its critic. We introduce Direction-Conditioned Policies (DCP), a method built around a small modification to CRL: DCP selects previously visited states as waypoints during online training and conditions the policy on their direction and distance in representation space. At deployment, DCP applies the same interface directly to the final goal, requiring neither waypoint selection nor planning. Across nine navigation and manipulation tasks, DCP attains higher final success rates than CRL on seven tasks and spends more time near the goal on seven. Controlled maze experiments further show that DCP captures shortest-path geometry more accurately and that the supplied direction causally influences the actor's behavior. We identify waypoint coverage and ranking as limits to exploration, and show that learned candidate generation improves goal reaching in two controlled mazes.

S K Swaminathan, Damiya Gondha, Theyanesh Eswaramoorthy Rajahkrishnan +1

Department of Computer Science and Engineering Indian Institute of Technology Kharagpur, India · Department of Mechanical Engineering Indian Institute of Technology Kharagpur, India

Reinforcement Learning Contrastive RL Goal-Conditioned RL

Oct 1, 2026cs.RO

InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation

We study test-time evolution for humanoid loco-manipulation: solving tasks that a controller was never trained for by repurposing its existing skills, improving from its own attempts, and retaining what it learns, without retraining. Our key insight is that a broad controller already holds much of the competence a new task needs, and that this competence becomes accessible through an interface between planning and control that is expressive enough to specify contact-rich, multi-stage interactions, yet executable and measurable enough that execution feedback can guide planning from experience. InterEvolve realizes this interface with two components. First, we develop an object-aware forward-backward (FB) behavioral foundation model, whose object residuals on a frozen body prior turn a new reward about the body or objects into loco-manipulation behavior at test time. Second, we specify tasks as reward programs: staged rewards with completion conditions and tunable constants. A large language model (LLM) agent revises the program structure in context, drawing on execution feedback and a skill library of verified programs, while a numerical optimizer tunes its constants. With every candidate verified across parallel simulation scenarios, the program explores new ways to induce, repurpose, and compose the controller's existing motor competence for the task at hand, and thus improves over iterations. Experiments show that human-designed rewards leave much of the FB model's loco-manipulation competence untapped, whereas the programs InterEvolve evolves release it, sometimes through novel strategies. It further produces behaviors for diverse tasks, complex scenes, and long-horizon compositions in simulation, and evolved skills run autonomously on a physical Unitree G1 from egocentric onboard perception.

Zhuo Lin, Sirui Xu, Liuyu Bian +2

University of Illinois Urbana-Champaign

Humanoid Loco-Manipulation Reinforcement Learning Long-Horizon Robotic Manipulation Large Language Model-Based Robot Planning Test-Time RL

Oct 1, 2026cs.LG

Faynt: Scaling and Optimizing Policies for Competitive Melee

We introduce Faynt, a family of 10M- and 75M-parameter Transformer policies for Super Smash Bros. Melee, each controlling all 26 characters with a single checkpoint. After reinforcement learning (RL), the 10M wins 240 of 244 same-character games (98.4%) against fourteen specialist and multi-character releases on their supported rosters, with a winning record against every release. These opponents retain 21- or 24-frame action delays; Faynt uses no added delay, and we have not isolated the effect of this difference. In a separate evaluation against a privately supplied zero-delay Slippi-AI model, the 10M wins all 68 games across two conditioning settings. We study architecture, optimization, scaling, and hyperparameter transfer to guide pretraining on approximately 840,000 human replays. Post-training combines rank- and outcome-based curricula, 75M-to-10M distillation, and RL restricted to Fox mirror matches. On the initial 152-game benchmark, the supervised 10M wins 69.7% of games, compared with 45.4% for the pretrained 75M, despite higher overall held-out controller-prediction loss. The weighted validation loss used for supervised checkpoint selection agrees with the win-rate ordering of all four pretrained and supervised policies. After supervised post-training, both models take less damage per minute, build larger early leads, and win more often after losing the first life. Optimized inference on recorded game states averages 5.2 ms per decision for the 10M and 8.7 ms for the 75M on an NVIDIA T4, excluding emulator execution and communication. We open-source the weights, both benchmark suites, and a platform for automated model tournaments.

Ali Janati, Nikita Kuzmin, Rohit Swamy +1

Frisson Labs

Reinforcement Learning Self-Play RL Transformer Imitation Learning

Oct 1, 2026cs.LG

Bellman Meets Lyapunov: Unsupervised Reinforcement Learning via Mastering Chaos

Reinforcement learning (RL) is a powerful paradigm for training agents, yet its success rests on domain expertise of human engineers who design informative reward signals for every new task. Unsupervised RL aims to reduce this engineering with intrinsic motivation (IM): reward signals that emerge from the agent environment interaction itself. Existing IM objectives, however, involve the selection of information variables, which re-introduces domain expertise the field has sought to eliminate. We introduce Forward CIP (F-CIP), an RL-native formulation of the Controllable Information Production (CIP) objective, which is defined by the system's dynamics alone and requires no such selection. We prove that F-CIP is compatible with RL and demonstrate its effectiveness with existing algorithms. Training agents with F-CIP results in unsupervised discovery of primitive behaviors such as balancing and maintaining controllability, which are essential for more complex robot behaviors. Paired with a simple forward-velocity reward, our method produces coordinated gaits such as hopping and running which otherwise require reward engineering to learn.

Tristan Shah, Wooyoung Chung, Volodomyr Makarenko +2

Texas Tech University · Purdue University

Reinforcement Learning Intrinsic Motivation