The paper introduces a steering agent that supervises large language model (LLM)-based agents and multi‑agent systems during execution to prevent goal drift. By observing the execution context, the steering agent decides when to intervene using either a low‑cost fast path or a high‑cost slow path, and both paths are jointly trained with reinforcement learning. Experiments on information‑seeking search tasks demonstrate that the steering agent improves performance and efficiency for both single‑agent and multi‑agent systems, generalizes to out‑of‑domain tasks, and allows controllable intervention frequency without retraining.
By Leyang Shen, Yaxin Gong, Yang Zhang, Chenxiao Fan, Jiakai Wang, Dong Wang, Yang Liu, Fuli Feng, Chun Kai Ling, Tat-Seng Chua
The paper introduces DSP: Dexterity from Supervised Play, a framework where a robot learns a play policy from minutes of human play via behavior cloning. The robot then autonomously uses this policy to gather hours of interaction data, training a world model that guides model‑predictive control during inference. The approach is validated on underactuated pushing and in‑hand pose reorientation tasks, outperforming hindsight behavior cloning by 69.6% in real‑world success rate and matching reinforcement learning performance with 490× less data.
By Botao He, Yiming Dou, Yiannis Aloimonos, Jitendra Malik, Akash Sharma, Ruoshi Liu
The paper introduces SLPR, a self‑learning framework that uses proximal policy optimization to optimize free‑form pipe routing in aeroengines. It incorporates a unified rule‑modeling system for obstacle detection, a potential energy table for rapid layout queries, and iteratively refines routing via agent‑environment interaction. Experiments show SLPR outperforms three baselines in pipe length, rule adherence, path complexity, and computational efficiency, and adapts quickly to changing design environments.
By Caicheng Wang, Zili Wang, Shuyou Zhang, Yongzhe Xiang, Zheyi Li, Jianrong Tan
The paper investigates how minimal decision-making relates to contextual probability using the Tug-of-War (TOW) principle for multi‑armed bandit problems. By fixing reward‑learning and information requirements, the authors show that minimizing relaxation of common event semantics can produce nonclassical probability, deriving the exact minimum semantic repair and conditions for selecting contextual statistics. They further develop a quantum Tug‑of‑War (QTOW) instrument that improves stationary decision performance, surpasses the KCBS noncontextuality bound, and provides a diagnostic for a corrected contextuality witness.
By Song-Ju Kim
Q($\gamma_H$) introduces a multi-horizon Q-learning framework that keeps separate discount-specific heads and blends them during control using an entropy-derived confidence signal. The adaptive mixing shifts from short to long horizons as policy confidence grows, allowing the agent to adjust its planning horizon without altering individual Bellman updates. Experiments on classical control and deep reinforcement‑learning tasks show the method excels on delayed‑credit problems compared to fixed‑discount baselines.
By Arwa Albelaihi, Abdulrahman Altahhan
The paper introduces a method for rational synthesis in stochastic multi‑agent systems with unknown dynamics, where each agent has a probabilistic temporal objective expressed in Probabilistic Linear Temporal Logic on Finite Traces. It formulates per‑agent rational synthesis as finding a strategy profile that satisfies the designated agent’s objective while also constituting a stable solution under chosen equilibrium concepts such as Nash or subgame‑perfect equilibrium. The authors provide conditional convergence guarantees for tabular Nash‑Q learning, certify satisfaction with finite‑sample confidence bounds, and evaluate strategic stability via approximate best‑response checking, validating the approach on eight general‑sum concurrent stochastic games.
By Omar Adalat, Edwin Hamel-De le Court, Francesco Belardinelli
The paper introduces Batch Cost Lookahead (BCL), a coordination‑free data‑loading strategy that buffers a bounded window of records per rank and repartitions them into fixed‑size packs with balanced estimated cost. BCL maintains the trainer’s record‑count contract without requiring cross‑rank communication or runtime changes, and in simulations it reduces per‑rank batch‑cost variability by up to 82.8%. Experiments on 13 NeMo‑RL SFT corpora across eight H200 GPUs show speedups ranging from 1.050× to 1.167× over pooled MFFD, with further gains up to 1.343× on multi‑node setups, while also identifying a memory‑feasibility boundary that limits outliers before balancing.
By Tao Huang, Chen Zhu, Antonio Alonso, Daming Li, Adarsh Goura, Jaideep Ray, Yitong Zhou
The paper explores the use of set membership estimation (SME) for uncertainty quantification in linear contextual bandits, contrasting it with the more common confidence‑interval approach. An SME‑based optimism‑in‑the‑face‑of‑uncertainty (OFU) algorithm is proposed, and a sublinear regret bound is proven under suitable assumptions. Simulations show that SME‑OFU performs favorably compared to existing methods.
By Haonan Xu, Yingying Li
RenoLayout is a new framework that unifies floorplan renovation and furniture layout for residential design. It takes an existing residential layout and a set of design points, then generates a Geometry-as-Code program that preserves immutable structural elements while optimizing furniture placement under design objectives and physical constraints. The authors also introduce a Design Map with 123 fine‑grained design points, a Difference‑Grounded Design‑Point Annotation method, and a three‑stage training pipeline (CPT–SFT–RL) to improve spatial reasoning, instruction following, and constraint compliance. Additionally, they present the RenoLayout Benchmark, a collection of 200 real‑world before‑after renovation cases with complete furniture layouts, marking the first benchmark for unified residential design.
By Jingguo Liu, Yuanqing Liu, Chong Zhang, Yue Yang
The paper introduces PACE, a method that enhances multimodal reinforcement learning with verifiable rewards (RLVR) by using pairwise token contrast and adaptive trajectory confidence. Pairwise token contrast evaluates each token under four teacher views—combining correct/incorrect hints with image presence or absence—to isolate image-dependent support, while adaptive trajectory confidence normalizes teacher predictions to guide credit redistribution. Experiments demonstrate that PACE outperforms existing approaches on multiple multimodal reasoning benchmarks, achieving top accuracy across three backbones and the Qwen3‑VL‑8B‑Instruct model.
By Yiran Zhang, Hanbo Huang, Jielan Li, Zhen Zhao, Shiyu Liang, Hongxia Hao
The paper introduces a method for training large language models to generate diverse, high‑quality scientific ideas grounded in input documents. By using a set‑level reinforcement learning objective that rewards ideas for being distinct from earlier ones, the authors train an 8B model that outperforms Claude Opus 4.8 and GPT‑6 Astra on a contamination‑free benchmark. Human experts rate the model’s idea sets as equally promising while containing slightly more unique ideas, and automated experiments show comparable executability.
By Manya Wadhwa, Pranav Narayanan Venkit, Hiroaki Hayashi, Yu Li, Tuhin Chakrabarty, Chien-Sheng Wu
The paper introduces OmniVerifier-C1, a graded visual verifier that decomposes complex prompts or environmental rules into fine‑grained components and assigns continuous pointwise scores, moving beyond binary True/False judgments. Experiments show that this structured, interpretable output improves generalization to out‑of‑domain multimodal critic tasks and outperforms binary verifiers in text‑to‑image scenarios. In rule‑governed generative world modeling, the authors demonstrate that a pairwise verification approach yields more accurate critiques than pointwise scoring, with empirical validation on tasks such as Maze and Sudoku.
By Bowei Liu, Xinchen Zhang, Xuhuan Li, Tianlin Zhang, Kaian Jiang, Jingwei Liu, Congyang Zhao, Yifei Qian, Junhong Liu, Zhiheng Li, Yujiu Yang, Yang Cai, Ling Yang
Writer‑R1 is a writing training pipeline that uses Inductive Rubric Generation (IRG) to automatically construct prompt‑specific rubrics for open‑ended writing tasks. The pipeline expands an open writing corpus, applies the generated rubrics during supervised training and reinforcement learning, and provides the rubrics to writers before generation. A compact model, RubricGen‑4B, can supply rubrics for new requests, and Writer‑R1 improves WritingBench and HelloBench scores over 4B and 8B backbones.
By Jihao Zhao, Shuaishuai Zu, Zhiyuan Ji, Chunlai Zhou, Biao Qin
The paper introduces Adaptive Data Scheduling (ADS), a dual-level framework that improves reinforcement learning for large language models by tailoring data sampling to both semantic clusters and the policy’s current capability. ADS first reorganizes training samples into semantic clusters and adjusts the inter-cluster sampling distribution to focus on clusters the policy can currently handle. Then, within each cluster, ADS selects prompts near the policy’s capability boundary to provide informative relative advantages. Experiments on three LLMs and seven reasoning benchmarks show ADS boosts average accuracy by 5.5% over Group Relative Policy Optimization (GRPO) and consistently benefits various RL objective designs.
By Zicheng Xu, Ruixuan Zhang, Yu-Neng Chuang, Xiuyi Lou, Hoang Anh Duy Le, Oren Gal, Alexander S. Szalay, Zhaozhuo Xu, Guanchu Wang, Vladimir Braverman
The paper introduces TextDART, a defect-aware reward model that evaluates individual text regions for semantic correctness and glyph quality, and SCANNER, a method that uses these regional judgments to provide spatially differentiated policy feedback. This approach addresses limitations of previous reinforcement learning methods that overlook glyph defects and obscure regional quality. Experiments on existing benchmarks and the new PolyText-Bench show state‑of‑the‑art performance in text accuracy and glyph fidelity.
By Sujie Hu, Huaisong Zhang, Pujing Yang, Qian Wang, Xiaomin Li, Xinhang Leng, Chen Li, Jing Lyu, Xiu Li, Ling Pan
The paper introduces ACT3D, a vision‑language model that learns to think and explore in 3D environments using only single‑image inputs and action‑based interactions. By fine‑tuning with trajectory supervision and then applying reinforcement learning driven by final‑answer correctness, ACT3D develops an exploration policy that adapts to the question and improves accuracy across six benchmarks, outperforming both proprietary and open‑source baselines while using fewer input frames. The study shows that outcome‑based RL can teach VLMs to decide when and how to acquire visual evidence, leading to better reasoning and spatial understanding.
By Yijiang Li, Bingyang Wang, Nuno Vasconcelos
SCOPE introduces a reinforcement learning approach for text‑to‑image prompt engineering that respects semantic spans such as objects, counts, attributes, and relations. By pooling policy log‑probability shifts within each semantic span, the method assigns a shared coefficient to token losses while using a single advantage per completion, eliminating the need for separate span rewards or extra image renders. Experiments on Qwen‑Image, SD3, and FLUX.1‑dev show average relative gains of 9.48% on GeneVal and 7.62% on T2I‑CompBench compared to vanilla GRPO, with ablation studies confirming the benefit of semantic boundaries and pooling.
By Yufei Jiao, Jinfa Huang, Zeyue Xue, Tianyu Xie, Yuexiao Ma, Rongrong Ji, Xiawu Zheng, Jiebo Luo
Routing Policy Optimization (RPO) introduces stochastic exploration directly within a visual generator’s attention mechanism, avoiding costly rollout-based optimization. By sampling key attention blocks at selected denoising levels and perturbing the transition velocity, RPO uses a Gaussian likelihood and a two‑rollout estimator to provide an advantage signal for learning. Experiments on text‑to‑image and text‑to‑video tasks demonstrate improved generation quality and nearly four times higher training efficiency compared to Flow‑GRPO‑style methods.
By Qiuhong Shen, Yiding Yang, Yuxin Zhang, Xinwei Huang, Minxuan Lin, Bo Liu, Hao Kang, Haotian Yang, Yizhi Wang, Chongyang Ma, Xinchao Wang
GRID‑Planner introduces a graph‑structured relational reward shaping framework for diffusion‑based autonomous driving planning. By integrating graph‑aware importance weighting, a time‑indexed relational reward surrogate, and an advantage‑weighted diffusion objective with on‑policy self‑distillation, it forms a rollout‑return‑advantage‑update loop that eliminates explicit likelihoods. Experiments on nuPlan and interPlan demonstrate significant reductions in collision rate, lane deviation, and jerk, along with a 3.4% improvement in overall planning score.
By Zhao Yang, Zezhong Qian, Yinan Shi, Longjun Liu, Yawei Jueluo
SceneReVis is a vision‑grounded, self‑reflective framework for 3D indoor scene synthesis that uses an iterative diagnose‑and‑act loop to detect and fix spatial conflicts such as collisions. It introduces SceneChain‑12k, a large dataset of causal construction trajectories generated via a reverse engineering pipeline, and employs a two‑stage training recipe that moves from supervised fine‑tuning to agentic reinforcement learning, turning the model into an active spatial planner. Experiments show that SceneReVis achieves state‑of‑the‑art performance in high‑fidelity generation and goal‑oriented optimization, with better transfer to underrepresented room categories and potential for out‑of‑distribution room types.
By Yang Zhao, Shizhao Sun, Meisheng Zhang, Yingdong Shi, Xubo Yang, Jiang Bian