The paper introduces Policy Gradient Penalty (PGP), a single‑loop policy‑space method that enforces convex occupancy‑measure constraints via quadratic‑penalty regularization. PGP constructs pseudo‑rewards to estimate gradients of the penalized objective and uses the classical Policy Gradient Theorem, establishing smoothness and global last‑iterate convergence guarantees for an ε‑optimal constrained entropy value with ε‑bounded constraint violation. The authors validate PGP with ablations on a grid‑world benchmark and demonstrate scalability on two challenging continuous‑control tasks.
By Florian Wolf, Ilyas Fatkhullin, Niao He
The paper introduces a curriculum reinforcement learning approach to overcome the cold‑start problem in prompt‑injection red‑teaming of frontier large language models. By training an attacker LLM sequentially against increasingly robust target models and ensuring partial success at each stage, the method achieves high attack success rates (93.8% against GPT‑5.6‑Luna and 45.0% against GPT‑5.6‑Terra) where prior RL methods fail. The attacker LLM also transfers its effectiveness to other frontier models it was not explicitly trained on.
By Chenlong Yin, Xiaolong Jin, Wei Zou, Yanting Wang, Jinyuan Jia
SIPO (Self‑Instructing Policy Optimization) unifies reinforcement learning with on‑policy self‑distillation by using a contrastive self‑teacher to generate token‑level credit signals. The method samples multiple rollouts per prompt, pairs each with a reference answer and its mistakes, and uses the difference in teacher log‑probabilities to provide dense feedback while still respecting the overall task reward. Experiments on reasoning and code‑generation benchmarks show that SIPO outperforms both RLVR and OPSD baselines without requiring an external teacher or extra generation steps.
By Zhenrui Yue, Huimin Zeng, Yueqi Wang, Yaokun Liu, Fengran Mo, Jinghan Zhang, Mung Yao Jia, Gyuseok Lee, Yang Zhang, Na Wei, Dong Wang
The paper introduces C-Instrument, a constitution‑grid data instrument designed to automate reinforcement learning (RL) data generation and hillclimbing. It also presents C‑LIM, a per‑cell learnability score that guides actions such as prune, densify, amend, or expand, and identifies dead‑weight data regions before training. The authors report that their method improves over‑refusal rates from 22.4% to 12.8% and reduces under‑refusal on adversarial attacks from 0.27 to 0.33, while lifting learning impact in a previously ineffective data region from 0.733 to 0.80.
By Xianling Zhang
AdaTutoRank introduces a setwise document reranker that uses Adaptive Tutoring Optimization (ATO) to provide graded supervision across nine rubric dimensions. By generating hint‑based silver labels, reinforcement rewards, and distillation cues tailored to each rollout’s quality, the method improves credit assignment for individual documents within a set. Experiments on ten benchmarks covering Retrieval‑Augmented Generation (RAG), deep research, and setwise evaluation show that AdaTutoRank achieves superior overall performance while reducing the number of retrieval calls.
By Kailin Jiang, Lei Liu, Jian Xi, Yangqi Chen, Hui Xu, Hongwei Zhao, Bin Li, Yu Lu, Haibo Shi
arXiv:2502.04492v3 Announce Type: replace
Abstract: The advancement of LLMs and their accessibility have triggered renewed interest in multi-agent reinforcement learning as robust and adaptive framew...
By Selim Furkan Tekin, Gaowen Liu, Ramana Rao Kompella, Ling Liu
arXiv:2609.36058v1 Announce Type: new
Abstract: Recent progress in reinforcement learning with verifiable rewards (RLVR) has highlighted the effectiveness of simple critic-free policy-gradient method...
By Hsiao-Ru Pan, Florent Draye, Bernhard Sch\"olkopf
arXiv:2609.36064v1 Announce Type: cross
Abstract: Foundation models are powerful generators, but many engineering domains require structured representations that general-purpose systems handle poorly...
By Sahand Rezaei-Shoshtari, Patryk Wozniczka, Shu Ishida, Gregg Streuber, Farnoosh Javadi, Jeffrey Landes, Angela Ju, Muhammad Azam, Bryan Lim, Johan Luttun, Indrajeet Haldar, Jonathan Shaw, Beatriz Guerra, Ivan Sosnovik, James Stoddart, Robert Giaquinto, Adam Gaier
arXiv:2609.36813v1 Announce Type: new
Abstract: Large language models (LLMs) exhibit strong general capabilities that mechanistic interpretability has attributed to sparse computational circuits. How...
By Chuanpu Liu, Miao Yu, Yikai Cai, Yuanhe Zhang, Zhenhong Zhou, Li Sun, Zuming Jiang, Yufei Guo
arXiv:2609.37065v1 Announce Type: new
Abstract: Decision-making under uncertainty often relies on predicted parameters, yet accurate prediction does not necessarily lead to good operational decisions...
By Jiahui Feng, Dafang Zhao, Zheng Chen, Zhengmao Li, Lingwei Zhu
Grounded Transitive RL (GTRL) is an offline goal‑conditioned reinforcement learning algorithm that improves upon divide‑and‑conquer value learning by grounding updates with a one‑step temporal‑difference (TD) target. By adding this TD target rather than replacing it, GTRL ensures every state‑goal pair receives an update and corrects bias from hindsight relabeling through reweighting based on reachability. The method was evaluated on nineteen OGBench tasks across stochastic, deterministic, and stitching environments, achieving the highest average success rate among compared approaches.
By Abdul Monaf Chowdhury, MD Sameer Iqbal Chowdhury, Shifat E Arman, Md Mehedi Hasan
PowerZooJax is a JAX-based benchmark suite designed for reinforcement learning in power system operation. It offers five constrained Markov decision process tasks covering generation, transmission, distribution, distributed energy resources, and data center microgrids. By implementing power flow, economic dispatch, market clearing, and device dynamics as JAX computation graphs, the entire training and evaluation loop runs on the GPU, yielding significant speedups over CPU-based simulations and enabling standardized evaluation of policy returns, safety violations, and out-of-distribution stress conditions.
By Zhanhua Pan, Xiao Liu, Zhilong Cao, Jianhong Wang, Dawei Qiu
CoRe introduces a co‑evolving reward framework to mitigate latent reward hacking in video diffusion models. By continuously refitting the latent‑reward model on the generator’s current samples and anchoring it to real‑video preferences, CoRe prevents the generator from drifting outside the reward model’s training support. Experiments on Wan2.1‑T2V‑1.3B demonstrate that CoRe improves generation quality over pretrained models and prior alignment methods while avoiding quality collapse.
By Zhaolong Su, Yujin Han, Feng Wang, Jameson Dong, Hins Hu, Difan Zou
The paper introduces BRIDGE, a bilevel optimization framework that jointly trains a large language model (LLM) and a retriever for agentic reinforcement learning (ARL). It demonstrates that adapting the retriever before the policy yields better rewards, and that BRIDGE outperforms existing methods on seven open‑domain QA benchmarks and medical QA tasks, achieving significant gains in accuracy and reasoning quality.
By Quan Xiao, Mingda Liu, Gaowen Liu, Katsuki Fujisawa, Tianyi Chen
The paper introduces TrustSwap, a counterfactual test that swaps or removes source reliability labels while keeping evidence text constant, to evaluate how retrieval‑augmented fact‑checking models respond across verdict, confidence, and search decisions. Experiments on untrained and RL‑trained models show that confidence and search largely follow labels, yet label changes can flip a significant portion of verdicts, especially in larger models. The authors propose trust‑swap augmentation (TSA) to mitigate this shortcut, demonstrating reduced verdict flip rates and maintained accuracy in several settings, though its effectiveness diminishes at larger model scales.
By Jianchang Su, Yiwei Yang, Wei Zhang
HorizonFlow is a hierarchical planner for offline goal-conditioned reinforcement learning that treats the planning horizon as an output rather than a fixed input. It uses a subgoal route planner and an action-prefix controller, both employing insertion-based generation and flow matching, to jointly generate continuous plan content and its length. The method leverages the partially generated plan to guide token insertion and to steer generation toward shorter plans, achieving superior performance on Maze2D, Multi2D, and OGBench benchmarks.
By JunHyeok Oh, Zian Jang, Byung-Jun Lee
The paper introduces FastRL, a reinforcement learning framework designed to enhance the efficiency of Group Relative Policy Optimization (GRPO) and its variants. FastRL employs an advantage-aware pruning strategy that retains high-advantage trajectories while preserving gradient diversity, and an adaptive rollout sampling mechanism that adjusts sampling scale during training based on historical pruning data. Experiments show that FastRL can be integrated into GRPO, DAPO, and GSPO, yielding a 2.07× speedup on Geometry3K and GeoQA8K-R1V and a 1.64% accuracy improvement on visual reasoning benchmarks.
By Jiahua Yang, Zhiwei Yang, Xianpeng Zhang, Dongyu Chen, Xing Chen, Tianhuang Su, Haonan Lu, Quanlong Guan, Kai Tang, Chuangchuang Wang
CoEM introduces a Commit-on-Evidence Memory system that learns when to compress source evidence into compact memory facts while preserving potentially useful excerpts verbatim in a pending set. The system uses a learned policy to decide whether to promote, retain, or discard each pending excerpt as new context arrives, and a frozen verifier ensures only supported facts are committed. Reinforcement learning trains this policy with step-level evidence rewards and final answer rewards, leading to consistent improvements in long-context reasoning, achieving 10.4–11.4 F1 points over the strongest baseline on 6,400-document inputs.
By Jingguang Li, Yebo Wu, Zuyi Guo, Kailang Ma, Xianjie Dai, Han Zheng, Benwang Chen, Li Li, Can Rong, Heye Huang
The paper introduces a Physics‑Informed Multi‑Agent Coordination framework that embeds calibrated BCMP queueing topologies into a decentralized multi‑agent reinforcement learning system for hospital patient flow. It formulates the problem as a Decentralized Partially Observable Markov Decision Process with coupled resource constraints, enabling departmental agents to negotiate patient routing and service scaling while exchanging localized action fingerprints to handle non‑stationarity. Empirical tests on MIMIC‑IV data show the approach reduces cumulative system delay compared to static Markovian models, heuristic dispatching, and independent multi‑agent baselines, all while preserving clinical safety constraints.
By Guoqing Zhang, Rafik Hadfi, Takayuki Ito
The paper introduces Watch-Think-Interact (WTI), a closed-loop framework for multi-question streaming video reasoning that maintains compact natural-language memory entries linked to video time ranges. WTI decides whether to answer, continue watching, or recall relevant past intervals for each question, avoiding replay of the full history. The authors build a large dataset, WTI-82K, and a training method, Stream-GDPO, achieving state‑of‑the‑art performance on StreamingBench and OVO-Bench.
By Ziheng Huang, Yicheng Bao, Xueheng Li, Zhenkun Gao, Bangwei Liu, Kunquan Li, Yuxiang Shen, Bangyan Li, Xuejiao Wang, Changbo Wang, Gaoqi He