EviRover is a perception agent that goes beyond a single glance by actively gathering information to resolve perceptual queries. The authors created two data generation pipelines, producing EviRover-SFT-5K and EviRover-RL-12K, and a human‑verified benchmark called EviLens with 688 instances across five perception categories. Trained with supervised fine‑tuning and agentic reinforcement learning, the 4B EviRover outperforms its backbone by an average of 30 points on EviLens and shows strong transfer to other benchmarks such as WebEyes and BrowseComp‑VL.
By Kaixuan Fan, Kaituo Feng, Tianshuo Peng, Yilei Jiang, Manyuan Zhang, Junke Wang, Xiangyu Yue
arXiv:2609.39021v1 Announce Type: new
Abstract: Reinforcement learning (RL) has substantially improved the reasoning ability of multimodal language models through verifiable rewards and increasingly...
By Haiying He, Xin Zheng, Shaoli Hu, Shijun Xiao, Xuanhe Liu, Bing Li, Harry Yang
arXiv:2609.38672v1 Announce Type: cross
Abstract: Beam-search-based test-time methods provide an effective way to improve large language model (LLM) performance on long-horizon generation by pruning...
By Qijia He, Yu Huang, Yuan Cheng, Yuxin Chen, Yingbin Liang
The paper introduces Length Self-Distillation (LSD) to address the length‑scaling tax (LST) that occurs during reinforcement‑learning post‑training, where models produce unnecessarily verbose responses to already‑solved prompts. LSD routes solved prompts to an on‑policy distillation process while keeping the original RL objective for unsolved prompts, using an exponential moving average of the online policy as its teacher. Experiments show LSD matches or surpasses RL performance while reducing LST from 19.0% to –3.7% on single‑turn reasoning and from 31.4% to 13.7% on multi‑turn agentic tasks, thereby maintaining concise responses on easy queries while still enabling exploration on difficult ones.
By Xu Wan, Wenyue Xu, Shengjie Zhao, Mingyang Sun
arXiv:2609.38863v1 Announce Type: cross
Abstract: The two-dimensional irregular knapsack problem in a fixed circular container is an important combinatorial optimization problem for maximizing materi...
By Zhongman Du, Huiming Zhang, Linlin Yang, Sheng Xu, Baochang Zhang
arXiv:2609.40137v1 Announce Type: cross
Abstract: We present Game-Guided Skill Discovery (GGSD), a framework that uses self-play in games to discover motor skills that are directly playable by humans...
By Seungeun Rho, Jeonghwan Kim, Xue Bin Peng, Sehoon Ha
arXiv:2609.38653v1 Announce Type: cross
Abstract: Recent advances in musculoskeletal modeling and reinforcement learning have enabled muscle-actuated agents to reproduce increasingly complex human mo...
By Merkourios Simos, Chengkun Li, Bianca Ziliotto, Alexander Mathis
arXiv:2609.38792v1 Announce Type: new
Abstract: We study training LLM judges from natural language feedback, especially for subjective tasks where the verdict depends strongly on which evaluation cri...
By Ilgee Hong, Changlong Yu, Zhenghao Xu, Xin Liu, Yuwei Zhang, Qin Lu, Bing Yin, Tuo Zhao
The paper investigates transferring specialized task-oriented behavior to a general language model without training or distillation. It applies two training‑free heterogeneous merging techniques—Intersection‑Merge (IM) and Activate‑Prune‑Merge (APM)—to project a specialist donor into the recipient’s parameter space and interpolate backbone weights. Experiments across embedding, reranking, reward modeling, and MoE code‑specialist tasks show that both methods improve the general model, demonstrating that simple parameter‑level merging can transfer capabilities across diverse specialist roles.
By Jiahe Fan, Si Chen, Yinghao Hou, Wenbo Xia, Ke Xu, Hong Xie, Enhong Chen
arXiv:2609.39868v1 Announce Type: new
Abstract: Exploratory reinforcement learning (RL) on an operating bus fleet is impractical,while policies trained only from historical data cannot acquire new ex...
By Yifan Zhang, Qifan Zhang, Liang Zheng
GrammarRL introduces a label‑free reinforcement learning approach that adapts language models to grammar constraints without annotated data. It optimizes two self‑supervised rewards—direct and reverse—using a Reinforce Leave‑One‑Out objective over grammar‑constrained rollouts, and regularizes toward a frozen base model. Experiments on sign‑language gloss translation, hierarchical text classification, and named entity recognition with Llama models show consistent gains over constrained greedy decoding and competitive performance to beam search while keeping inference cost low.
By Gabriele Tuccio, Antonino Furnari, Aldo Gangemi, Misael Mongiov\`{\i}
arXiv:2609.40035v1 Announce Type: new
Abstract: The policy-gradient theorem gives the exact gradient under the current policy, but finite on-policy samples may miss rare high-return trajectories. We...
By Junyu Lu, Shichao Weng, Zhiqiang Wang, Haojie Luo, Jingfan Zhang, Yuhua Zhou, Cheng Du, Yuzhuo Zhang, Xi Li, Jinwei Du, Tiancheng Feng, Chuan Xiao, Shuyuan Zheng
arXiv:2510.01581v2 Announce Type: replace-cross
Abstract: Recent thinking models are capable of solving complex reasoning tasks by scaling test-time compute, but this scaling should be allocated in l...
By Joykirat Singh, Justin Chih-Yao Chen, Archiki Prasad, Elias Stengel-Eskin, Akshay Nambi, Mohit Bansal
arXiv:2605.15692v2 Announce Type: replace-cross
Abstract: We study episodic reinforcement learning with fixed reward and transition functions, but with episode-dependent admissible action sets that a...
By Zijun Chen, Zihan Zhang
arXiv:2605.10194v2 Announce Type: replace
Abstract: On-policy self-distillation (OPSD) uses a model as its own teacher under privileged context, providing token-level supervision on the model's own r...
By Jiaxuan Wang, Xuan Ouyang, Zhiyu Chen, Yulan Hu, Lan-Zhe Guo
StateTree is a reinforcement learning approach that improves long‑term dialogue reasoning by building a tree‑structured auxiliary task from limited dialogue data. The method embeds key‑value records across multiple sessions into a binary tree, requiring the model to traverse from root to leaf, retrieve records, compare timestamps, and identify a target question among distractors. Curriculum RL training increases tree depth, and a compositional variant trains the model to combine partial reasoning fragments, enabling cross‑session retrieval, temporal reasoning, knowledge updates, and multi‑hop reasoning while generalizing from 10K‑token to 128K‑token contexts.
By Naen Xu, Wanqing Cui, Yibo Hu, Shixin Hong, Hengyu An, Meiguang Jin, Junfeng Ma, Tianyu Du
arXiv:2606.15911v2 Announce Type: replace
Abstract: This paper focuses on automatically generating informative ad descriptions in sponsored search. Unlike ad titles which are usually optimized to att...
By Penghui Wei, Jiayu Wu, Chao Ye, Zhi Guo, Shuanglong Li, Lin Liu
The paper introduces Trajectory-guided Joint Policy Optimization (TJPO), a reinforcement‑learning framework that explicitly encourages diversity in the trajectories of large‑language‑model agents. By defining task‑specific trajectory descriptors, TJPO measures and optimizes diversity as a set‑level function, avoiding population‑based training. Experiments on Sokoban and ALFWorld demonstrate that TJPO yields diverse, interpretable behaviors while preserving strong task performance.
By Huaiyu Fu, Heng Cao, Hao Wang, Jian Ya, Tao Chen
LexReward is a taxonomy-driven reward framework designed for legal language models, evaluating responses across three dimensions: Style (lexical and syntactic quality), Element (legal subjects, facts, statutes, and decisions), and Chain (order, completeness, correctness, and non-redundancy of reasoning). The framework introduces rubrics that define evaluation criteria and quality levels for each dimension, generating pairwise preference data used for Direct Preference Optimization (DPO) and reward-model training. Experiments demonstrate that rubric-based rewards effectively differentiate legal responses of varying quality, and that DPO training improves performance across all dimensions; the resulting reward models (LexRM) enable reinforcement learning to enhance policy performance in each specific dimension without needing reference answers.
By Yida Cai, Xin Dai, Bingxiang He, Huiyuan Xie, Yuxiao Ye, Zhenghao Liu, Yang Bai, Zhiyuan Liu
The paper introduces TacEx, a tactile‑curiosity framework that guides reinforcement learning agents to explore contact dynamics by focusing epistemic uncertainty on the tactile channel. By anchoring curiosity to touch, robots learn to manipulate and grasp objects without task rewards or demonstrations, generating an interaction‑dense dataset that supports offline pick‑and‑place policy learning. TacEx also enhances vision‑language‑action models through post‑training, improving downstream performance while remaining sample‑efficient.
By Klemens Iten, Alexander Proshkin, Bhavya Sukhija, Stelian Coros, Andreas Krause, Pieter Abbeel, Carmelo Sferrazza