The paper presents a modeling and simulation framework to study reinforcement‑learning control of connected and automated vehicle (CAV) platoon joining maneuvers in mixed traffic. It evaluates Deep Q‑Network, Double Deep Q‑Network, and Proximal Policy Optimization algorithms, finding that PPO achieves a 98 % joining success rate with less than 1 % collisions by incorporating risk penalties, though it requires more decision steps. An external safety controller can prevent collisions but may reduce joining efficiency, highlighting a trade‑off between safety, effectiveness, and decision speed.
Behavior2Trip introduces a new task—Behavior‑Aware Travel Planning—where user preferences are inferred directly from past behavior trajectories rather than explicit instructions. The benchmark contains 11,400 Chinese travel‑planning instances, each with an average of 39.8 past behaviors across 14 attributes and 5 preference dimensions. A reinforcement‑learning agent, B2T‑Agent, leveraging behavior trajectories, external retrieval tools, and internal memory, outperforms strong baselines such as GPT‑4.1 on this challenging dataset.
The paper introduces Instruct-to-Act, a system that decouples high‑level planning from low‑latency control by combining a vision‑language model (VLM) planner with a world‑model controller. The VLM generates sparse, high‑level text instructions, while the controller executes them autonomously at high frequency. Experiments across seven embodied environments, including multi‑agent settings, show that this approach outperforms both controller‑only and direct VLM action‑generation methods, maintains fast control, and allows swapping in different pretrained VLM planners without fine‑tuning.
The paper introduces a new approach to safety in contextual bandits with continuous actions by enforcing high‑probability constraints on the realized cost rather than on its expectation. It proposes the High‑Probability Constrained UCB algorithm, which balances optimistic reward exploration with pessimistic safety estimation. The authors provide theoretical regret guarantees for linear models and extend the analysis to general function classes, demonstrating experimentally that realized‑cost constraints significantly reduce safety violations compared to expected‑cost baselines.
The paper introduces Regularized Latent Dynamics Prediction (RLDP), a method that adds orthogonality regularization to self‑supervised next‑state prediction in latent space. RLDP maintains feature diversity, matching or surpassing complex representation learning approaches for zero‑shot reinforcement learning. It also performs robustly in low‑coverage data settings where prior methods fail.
By Pranaya Jajoo, Harshit Sikchi, Siddhant Agarwal, Amy Zhang, Scott Niekum, Martha White
The paper investigates whether large language models (LLMs) exhibit language‑specific skill differences by having two identical model instances play a text‑based game in different languages. Using a multilingual extension of TextArena, the authors evaluate three open‑weight models across eight languages and six games, finding that the same model can show markedly different performance—varying win–loss margins, invalid actions, and strategic choices—depending on the language interface. Analyses pinpoint language‑specific failures in spatial reasoning, card‑conditioned decisions, and optimal move selection, and demonstrate that adjusting the intermediate reasoning language can recover much of the lost performance.
By Bobby Cheng, Adam Gaber, Zhengyuan Liu, Catherine Arnett, Omer Goldman, Cheston Tan, Leshem Choshen
The paper reviews On‑Policy Self‑Distillation (OPSD), a method where a language model learns from its own generations using privileged information such as reference solutions or plans, eliminating the need for a larger teacher model. It identifies a key failure mode—collapse, where the model’s reasoning paths narrow progressively—and analyzes it through three levers: signal application, privileged information, and teacher dynamics. The review focuses on mathematical reasoning, offering a unified vocabulary and distinguishing settled facts from ongoing debates.
By Justin Robert, Raheel Qader
The paper introduces V‑Rubrics, a reinforcement‑learning framework that evaluates vision‑language model responses by breaking them into atomic propositions and scoring them on Visual Faithfulness, Reasoning Consistency, and Instruction Following. Using a fine‑tuned Qwen3‑VL‑8B‑Instruct model and a newly created 50K‑example V‑Rubrics dataset, the authors demonstrate that rubric‑based GRPO outperforms both a shared SFT baseline and an answer‑only GRPO, especially on knowledge‑oriented and visually grounded reasoning tasks.
By Shulin Tian, Minglun Li, Yuhao Dong, Hao Ding, Jiarui Yao, Haiwen Diao, Jingkang Yang, Hongyuan Zhu, Ziwei Liu
The paper introduces a novel framework that combines vision‑language model (VLM) generated preferences with the Plackett‑Luce (PL) ranking model for reward learning in reinforcement learning. Unlike traditional pairwise Bradley‑Terry approaches, the PL formulation allows listwise rankings of multiple candidates, enabling the use of different ranking sizes (K = 3, 4, 5). Experiments on Meta‑World manipulation tasks show that PL‑based reward models train robotic policies as effectively as, or better than, pairwise, K‑wise, and RL‑VLM‑F baselines, achieving up to an 86% mean final success rate and matching the Oracle baseline on the Drawer Open task.
By Srivalli Katkuri, Maxwell Kawada, Juan Wachs
VGA‑BenchV2 is an expanded, human‑aligned benchmark and optimization framework that jointly evaluates video generation quality and aesthetic value. It builds on the original VGA‑Bench taxonomy, adding 52 sub‑dimensions and 1,016 curated prompts to generate over 60,000 videos from 12 mainstream models. The benchmark significantly enlarges human supervision with 36,000 task‑level annotations and introduces a hybrid evaluator (VAQA‑Net, VTag‑Net, VGQA‑Net) that aligns well with human judgments and can be used as a reward model for reinforcement‑learning fine‑tuning.
By Longteng Jiang, DanDan Zheng, Qianqian Qiao, Heng Huang, Huaye Wang, Yihang Bo, Bao Peng, Jingdong Chen, Jun Zhou, Xin Jin
The paper introduces COMLLM, a generative framework that combines Group Relative Policy Optimization with a Look‑Ahead Collaborative Simulation to enable multi‑turn reasoning for task offloading in Mobile Edge Computing. By performing multi‑step Monte Carlo rollouts that jointly model server queue dynamics, COMLLM incorporates long‑term system evolution into its reward design, achieving near‑optimal latency and improved load‑balancing fairness. The framework demonstrates zero‑shot scalability to larger network topologies, outperforming supervised fine‑tuning, deep reinforcement learning, and heuristic baselines without requiring retraining.
By Ning Yang, Chuangxin Cheng, Haijun Zhang
AutoVerifier is a residual‑guided, non‑parametric optimization framework designed to improve reference‑based answer verification. It learns verifier inductive biases from recurring errors, records them as rule cards, and promotes them to code modules or prompt guidance only after replay validation ensures no regressions. Experiments on four verifier benchmarks show that AutoVerifier surpasses state‑of‑the‑art verifiers by a large margin.
By Zebei Zhao, Zhihao Shi, Minqi Shi
ResMerge is a new framework for merging large language models trained via reinforcement learning. It separates each model’s task vector into a leading spectral head and a residual component, finding that both parts contain valuable behavior knowledge but behave differently during merging. The method builds a stable residual backbone using Spherical Residual Consensus Adaptation and then adds a lightweight head correction module that activates only when experts agree, leading to better preservation of expert capabilities compared to existing merging baselines.
By Yandu Sun, Zhiyan Hou, Hongyan An, Weizhen Wang, Haokai Ma, Yuheng Jia, Junfeng Fang, Haiyun Guo, Jinqiao Wang
MMEmb-R1 is a multimodal embedding framework that enhances reasoning by treating it as a latent variable and selecting beneficial reasoning paths through pair-aware selection and counterfactual intervention. It uses reinforcement learning to invoke reasoning only when necessary, reducing unnecessary computation and latency. On the MMEB-V2 benchmark, MMEmb-R1 achieves a state‑of‑the‑art score of 71.2 with just 4 B parameters.
By Yuchi Wang, Dingkang Yang, Haiyang Yu, Weikang Bian, Jiefeng Long, Xiao Liang, Chao Feng, Hongsheng Li
The paper introduces pFedMARL, a federated learning framework that uses multi‑agent reinforcement learning with TD3 to dynamically adjust client contributions and personalize models. It applies a server‑side agent to optimize global aggregation and client‑side agents to balance global and local updates, eliminating the need for pre‑training. Experiments on a semi‑supervised audio spectrogram transformer show that pFedMARL outperforms or matches FedAvg, Ditto, and local training across various non‑IID settings and against adversarial clients, improving accuracy, robustness, and fairness.
By Rene Glitza, Luca Becker, Rainer Martin
The paper introduces STDSH-MARL, a multi-agent deep reinforcement learning framework that uses a dual-stage hypergraph attention mechanism to capture spatio-temporal dependencies in corridor traffic signal control. It employs a hybrid discrete action space to jointly set signal phase configurations and green durations, allowing more adaptive timing. Experiments on a corridor network show that STDSH-MARL outperforms state‑of‑the‑art baselines, notably reducing tram waiting times while balancing overall network efficiency, tram priority, and bus service quality.
By Xiaocai Zhang, Neema Nassir, Milad Haghani
The paper introduces Prefix Sliding, a method that discards intermediate reasoning tokens during test-time scaling of language models, keeping only the prefix with key instructions and the most recent few thousand tokens. This approach limits memory usage regardless of reasoning length, enabling efficient long-horizon scaling. Experiments show that without training, Prefix Sliding can triple inference speed while preserving performance, and with reinforcement learning it can further improve results on very long reasoning traces.
By Niklas Muennighoff, Zhengyang Wang, Zeyi Chen, Weijia Shi, Binyuan Hui, John Yang, Dapeng Jiang, Mika Senghaas, Fares Obeid, Johannes Hagemann, Sami Jaghouar, Ludwig Schmidt, Percy Liang, Jason Wei, Andrew Y. Ng, Luke Zettlemoyer, Yejin Choi, Mike Lewis
The paper introduces Golden-GRPO Injection (GRIN), a three-stage self‑learning framework that uses a mixed‑policy reinforcement learning algorithm to inject knowledge into large language models. GRIN injects a golden answer to provide learning signals even when on‑policy rollouts fail on novel facts, and is evaluated on two new document‑level benchmarks—Blank and Counter—that test novel acquisition and counterfactual overwrite. Experiments show that mixed‑policy RL enables knowledge absorption beyond what supervised fine‑tuning can achieve, with GRIN outperforming SFT and other RL baselines on harder question types while matching them on basic fact recall.
By Zhibo Hou, Fan Zhao, Zhiyu An, Wan Du
PointRL introduces a verifiable reinforcement learning framework that learns point-level vision‑language grounding from heterogeneous annotation evidence such as bounding boxes, masks, and instance labels. The method converts these annotations into pointing instructions while preserving target supports, instance membership, and set constraints as hidden verifier evidence, which a deterministic checker uses to score predictions. Evaluation on PointArena shows that PointRL improves Qwen3.5‑4B’s accuracy from 56.11% to 65.58%, and similar gains are observed on RoboSpatial, BLINK, and Ref‑Adv benchmarks.
By Jingyang Su, Pu Cao, Xiuze Jin, Longyue Zhang, Qing Song, Lu Yang
ShuttleArena is a physics‑based singles badminton self‑play environment that integrates continuous shuttle flight, player interception, structured shot generation, and post‑shot recovery. The policy employs role‑conditioned outputs, allowing interpretable tactical probes through masked interception choices for receivers and factorized hitter actions over shot azimuth, elevation, speed, and recovery target. Evaluation with frozen checkpoints, controlled tactical probes, recovery ablations, qualitative rollouts, and a human‑data sanity check demonstrates competitive performance and reveals that learned recovery behavior is critically important for success.
By Peize Ding