RL-ARC introduces a calibration-aware training framework that uses reasoning confidence to guide the calibration of answer confidence in large reasoning models. By applying reasoning-guided regularization for correct predictions and an overconfidence penalty for incorrect ones, RL-ARC improves calibration across both in-distribution and out-of-distribution settings. The approach maintains reasoning performance while enabling models to adaptively estimate confidence for each question.
The study compares supervised fine-tuning (SFT) and reinforcement learning (RL) as teacher training methods for on-policy distillation (OPD) in a single-teacher setting. Using Qwen3.5-9B models across Agentic, Reasoning, and Perception domains, RL-trained teachers consistently produce stronger students and better recovery of performance gains than SFT-trained teachers. The authors attribute this advantage to RL teachers’ smaller parameter shifts from the shared initialization, which facilitates learning through OPD.
By Jingyuan Huang, Zuming Huang, Yucheng Shi, Zhongzhi Li, Xiaoming Zhai, Wei Chu, Ninghao Liu
arXiv:2610.09665v1 Announce Type: new
Abstract: On-policy post-training can improve large language models by learning from their own trajectories, but requires costly rollout generation. We ask wheth...
By Tianle Wang, Jiayu Liu, Ruizhi Zhao, Ning Miao
The paper investigates how randomized exploration in bandit learning can be compromised when an adversarial agent gains access to another agent’s exploratory signals. In a two‑player strongly monotone setting, the deceptive agent couples its actions with the leaked information to steer the learning dynamics toward a new steady state, termed the deceptive Nash equilibrium (DNE). The authors prove that the deceptive bandit learning dynamics converge to an arbitrarily small neighborhood of the DNE while maintaining optimal convergence rates, even under relaxed smoothness assumptions, and illustrate the impact on costs in a resource‑allocation game.
By Michael Tang, Mahmoud Abdelgalil, Jorge I. Poveda
The paper introduces temporally interpretable differentiable decision trees (DDTs) that incorporate action chunking to better align tree-based decision making with multi‑timestep planning. Two new policy‑gradient algorithms and an information‑theoretic tree‑restructuring method are proposed to enable efficient training of these trees. Experiments on four simulation environments show that warm‑starting action‑chunked DDTs from a distilled policy yields trees that match neural network performance in three domains while using up to 80% fewer parameters.
By Eisuke Hirota, Aarav Sane, Rohan Paleja
The paper introduces Continual Graph Multi-Agent Reinforcement Learning (CGMARL), a framework that maps sequences of tasks into attributed graphs to capture task-specific structures such as network topologies or formation arrangements. It presents the Graph-based Formation (GRAFO) benchmark to illustrate forgetting in graph-based CMARL settings and proposes the Frozen Graph Encoder (FROG) to preserve past structural information using a frozen graph backbone. Experiments on GRAFO demonstrate that combining FROG with existing continual learning methods significantly improves performance across multiple CGMARL scenarios.
By Tommaso Marzi, Ahmed Hendawy, Jan Peters, Carlo D'Eramo, Andrea Cini, Cesare Alippi
Continuous-Utility Direct Preference Optimization (CU‑DPO) replaces binary preference labels with continuous scores to better capture fine‑grained reasoning quality. The framework aligns language models to multiple prompt‑based cognitive strategies, proving a θ(K log K) sample‑complexity improvement and convergence to an entropy‑regularized utility‑maximizing policy. A two‑stage pipeline—strategy selection via best‑vs‑all comparisons and execution refinement using margin‑stratified pairs—yields significant gains on mathematical reasoning, code generation, and causal reasoning benchmarks.
By Muhammad Ahmed Mohsin, Muhammad Umer, Emily Fox
SkillRL is a framework that enhances large language model agents by automatically discovering and evolving skills from raw experience. It builds a hierarchical skill library called SkillBank, uses an adaptive retrieval strategy for heuristics, and allows the skill library to co‑evolve with the agent’s policy during reinforcement learning. These techniques reduce token usage and improve reasoning, achieving state‑of‑the‑art results on ALFWorld, WebShop, and seven search‑augmented tasks, outperforming baselines by 15.3% and remaining robust as task complexity grows.
By Peng Xia, Jianwen Chen, Hanyang Wang, Jiaqi Liu, Kaide Zeng, Yu Wang, Siwei Han, Yiyang Zhou, Xujiang Zhao, Haifeng Chen, Zeyu Zheng, Cihang Xie, Huaxiu Yao
The paper introduces BehaviorTrace, an evaluation framework for assessing training‑data attribution in online reinforcement‑learning fine‑tuning of language models. Using a planted behavior on Qwen2.5‑1.5B, the authors show that many attribution signals are confounded by factors such as gradient magnitude and model fluency, and that only the gradient of trigger tokens consistently aligns with the true behavior across seeds. They provide a checklist for evaluating attribution methods and test existing estimators without proposing a new one.
By Amit Nautiyal
The paper introduces Exploration-Distillation (ExpDis), a method that separates exploration from optimization in reinforcement learning with verifiable rewards (RLVR) for language models. ExpDis trains explorer policies with a novelty bonus, filters their outputs for correctness, and distills the results into a student policy that is optimized without the novelty bonus. This iterative process enables aggressive exploration while preserving model quality, outperforming DAPO on multiple mathematical reasoning benchmarks and improving pass@$k$ scaling.
By Saif Punjwani, Micah Goldblum
TiTok is an audio‑visual large language model designed for multi‑segment temporal grounding in untrimmed videos. It introduces Time Token Interleaving (TTI) to align temporal perception with prediction, and uses decoupled reinforcement‑learning rewards (global, local, count, precision, format) optimized via GDPO. Evaluated on a new UnAV‑100 protocol with the CountF1 metric, TiTok achieves state‑of‑the‑art results (65.7 mIoU, 0.58 CountF1).
By Eunji Shin, Dahyun Choi, Seungyeon Jo, Yejin Hong, Jiyoung Lee
HydroSphere is a governed, data‑driven framework that enables real‑time monitoring, forecasting, treatment optimization, and fault recovery for wastewater infrastructure. It combines a hybrid TCN‑LSTM model for multi‑step water‑quality forecasting, a PPO‑based adaptive dosage optimization module, and a SHADE autoencoder for anomaly detection, all evaluated on 2.82 million measurements from 1940 to 2023. The system incorporates safety bounds, human oversight, and tiered governance to promote safer, more adaptive water management aligned with UN Sustainable Development Goals 6 and 13.
By Prabu, Fancy C, Suresh A, Srini Ramaswamy
LASER is a new offline reinforcement learning algorithm that performs latent-space adjoint matching with entropy regularization to avoid policy collapse and critic exploitation. It uses expressive flow policies without backpropagation through time and applies a single set of hyperparameters across all tasks. Experiments on 40 OGBench tasks show LASER outperforms baselines, achieving state‑of‑the‑art results with robust, task‑agnostic performance.
By Songyuan Zhang, Oswin So, Eric Yang Yu, Matthew Cleaveland, Peter Crowley-Dolen, Chuchu Fan
TAP (Trajectory‑Anchored Pruning) is a structural pruning framework designed for reinforcement‑learning‑trained agents that addresses the shortcomings of existing pruning methods on long‑horizon tasks. It couples iterative channel pruning with on‑policy recovery, anchoring student interactions to teacher trajectories and re‑scoring channels using gradients from the recovery objective. Experiments show that removing 60% of FFN channels preserves 99.2% and 88.0% of task success rates on ALFWorld and WebShop while cutting GPU time per successful task by about 22% and 17%.
By Yuanzhe Li, Pengxin Wang, Yuxin Ren, Jianing Deng, Jingtong Hu, Song Wang, Jingdi Chen, Huanrui Yang
The paper introduces Convex-Concave Reinforcement Learning (CCRL), showing that the exact per‑iteration objective in policy learning can be expressed as a difference‑of‑convex (DC) program in log‑density‑ratio coordinates. This formulation unifies existing methods such as CPI, NPG, TRPO, and AWR as special cases and enables a multi‑step axis that couples consecutive decisions. Using sequential convex programming, the authors provide convergence guarantees and demonstrate that CCRL outperforms or matches PPO on diagnostic MDPs, classic control tasks, and a stochastic mid‑horizon healthcare domain, achieving faster convergence and higher training‑curve area.
By Shripad V. Deshmukh, Yaswanth Chittepu, Dhawal Gupta, Philip Thomas, Scott Niekum
VideoEvolve is a self‑evolving framework for long‑video understanding that jointly updates both memory and retrieval mechanisms. It starts with a coarse low‑frame‑rate overview and uses a Memory Evolver to selectively augment memory while a Retrieval Evolver adapts how information is fetched from that memory. The two components are alternately refined via Agentic Reinforcement Learning, guided by Bottleneck‑Aware and Capability‑Aware Evolution Feedback to focus on the most limiting aspect and to prevent over‑specialization, respectively. Experiments on several long‑video benchmarks show that this approach improves performance over static systems.
By Yongchao Xu, Bowen Ye, Jiefeng Gan, Junkai Ma, Wenzhao Li, Sen Tao, Yi Wei, Jiawei Liu
The paper identifies an informational curse of horizon in goal-conditioned policy learning, showing that longer goal relabeling horizons during training can hurt policy generalization and performance. Controlled experiments with oracle planners separate training horizons from test horizons, revealing that behavioral cloning policies degrade with longer horizons while reinforcement learning mitigates this effect. The authors attribute this to reduced conditional mutual information between actions and hindsight-relabeled goals, and demonstrate that distilling short-horizon policy Jacobians into long-horizon policies improves performance, especially in combinatorial manipulation tasks.
By John L. Zhou, Yuxuan Dong, Jonathan C. Kao
EvoSignal is an LLM‑guided evolutionary framework that treats traffic signal control as a modular program design problem, separating traffic feature extraction, local phase prioritization, and optional network‑based priority adjustment. Starting from established strategies, it iteratively improves programs using congestion and signal operation feedback, ultimately producing control rules that operate without online LLM inference. In simulations on two real‑world road networks, the best EvoSignal program reduced waiting time by 16.8–49.2% compared to 20 conventional, reinforcement‑learning, and LLM‑based baselines, and a travel‑time/queue‑length prioritizing program consistently outperformed all baselines across multiple scenarios.
By Leizhen Wang, Peibo Duan, Zhenlin Qin, Yancheng Ling, Jian Xu, Yue Wang, Hao Wang, Zhenliang Ma
The paper introduces COPC, a Coupled Off-Policy Correction method for asynchronous reinforcement learning of large language models. COPC coordinates policy-side and advantage-side corrections by combining token-level ratio masking with two-sided clipped-ratio weighting of TD residuals, addressing both policy mismatch and advantage staleness. Experiments show COPC outperforms existing asynchronous baselines on tool-integrated mathematical reasoning and search tasks, while maintaining training stability and minimal overhead.
By Zicheng Hu, Zhijian Zhou, Xuan Zhang, Yuchen Liu, Cheng Chen, Yuan Li, Qi Gu, Yan Feng, Hongyan Hao, Chao Qu
RollVerify is a lightweight reinforcement learning framework that addresses the trade‑off between efficiency and accuracy in long‑tail rollout settings. It introduces an off‑policy shift metric (OPS) to quantify deviation in partially generated trajectories and uses sequence‑level and token‑level verification to truncate invalid suffixes before training. Experiments on mathematical reasoning tasks show that RollVerify matches on‑policy accuracy while cutting training costs, with preliminary evidence from code‑generation tasks.
By Yongqiang Yao, Jinru Tan, Kaihuan Liang, Zixin Yin, Yazhe Niu, Ruihao Gong, Dahua Lin, Ningyi Xu