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
BVR Sim is an open‑source, Gymnasium‑style environment for heterogeneous air‑combat reinforcement learning, supporting multiple JSBSim aircraft models (F‑15, F‑16, F/A‑18, F‑22) with configurable weapons, sensors, and opponents. It offers a unified tactical action interface, interchangeable Python and accelerated C++ backends, entity‑oriented observations, compositional rewards, scripted opponents, replay and visualization, and adapters for multi‑agent learning frameworks. At a 0.4‑second decision interval, the C++ backend achieves 104 simulated seconds per wall‑clock second in 1‑vs‑1 and remains practical through 10‑vs‑10 scenarios, and a policy trained on the F‑16 transfers to four unseen aircraft with a 45.5% mean win rate after controller adaptation.
By Haocheng Sun (Beijing University of Posts,Telecommunications), Mulai Tan (Air Force Engineering University)
AdaVDR is an adaptive video deep research agent that selects and reflects on tool usage based on the task and the model’s capabilities. It constructs a specialized data pipeline to generate high‑quality QA pairs and uses model‑conditioned filtering to remove unnecessary tool calls. The agent is trained with supervised fine‑tuning and reinforcement learning, achieving top performance on the VDR‑EE benchmark and significant gains on VideoDR.
By Xintong Zhang, Xiaomeng Fan, Shilin Yan, Ekko He, Zicheng Liu, Zijian Zou, Guannan Zhang, Yuwei Wu, Zhi Gao, Hongwei Xue
The paper introduces R4P, a reasoning‑based supervision method for software agents that eliminates the need for test execution by using a group‑wise training objective to verify multiple patches simultaneously. R4P achieves 72.2% accuracy on the SWE‑bench patch verification task, matching proprietary models, and enables the creation of an execution‑free scaffold called Mini‑SE. Mini‑SE, trained purely with reinforcement learning via R4P, improves Pass@1 from 26.2% to 32.8% over the baseline Qwen3‑32B, demonstrating R4P’s practical utility and scalable performance.
By Junjielong Xu, Boyin Tan, Xiaoyuan Liu, Chao Peng, Pengfei Gao, Pinjia He
TailSFT is a simple modification to supervised fine‑tuning that filters out already well‑modeled sequences, concentrating learning on the tail of the data distribution. On the OLMo‑3 7B model, this approach improves pass@16 performance on math and coding tasks by up to 17% absolute and yields up to 4% absolute gains in subsequent GRPO reinforcement‑learning runs, with only minimal computational overhead. The authors also provide a lightweight diagnostic to identify settings where TailSFT is most beneficial and argue for a stage‑aware development strategy that evaluates intermediate checkpoints by their support for later training.
By Sadhika Malladi, Samy Jelassi, Dylan Foster, Jordan T. Ash, Akshay Krishnamurthy
The paper introduces a method to reduce sycophancy in large language models by using the Bayesian Truth Serum (BTS) as a reward signal in Group Relative Policy Optimization (GRPO). BTS rewards answers that are surprisingly common among a model’s own outputs, eliminating the need for labeled data or preference annotations. Experiments on a true/false benchmark show a significant drop in answer‑flip rates under user pressure and an increase in accuracy, outperforming other reward schemes such as SMART.
By Serhii Mytsyk, Yiming Zhang, Vikram Krishnamurthy
The paper introduces Multi-Objective In-context Knowledge Editing (MO‑IKE), a reinforcement learning framework that treats prompt construction for knowledge editing as a constrained Markov decision process. MO‑IKE jointly optimizes three competing objectives—reliability, generality, and specificity—by training a dynamic retriever to balance these goals and produce globally coherent prompts. Experiments on Llama‑3.2 show that MO‑IKE raises edit success from 85.0 % to 92.0 %, improves paraphrase consistency from 77 % to 79 %, and boosts retention rate by 23 % compared to earlier RL‑based methods.
By Xuzhong Wang, Maiqi Jiang, Tejal Nair, Girija Bhusal, Yanfu Zhang, Haipeng Chen
The paper introduces T$^3$RL, a tool‑verification framework for test‑time reinforcement learning (TTRL) that mitigates the false‑popular failure mode by using external tool evidence to upweight verified rollouts during voting. By grounding pseudo‑label construction in verified evidence, T$^3$RL produces more reliable pseudo‑labels and improves performance over standard TTRL on math benchmarks such as MATH‑500, AMC, and AIME 2024. The approach positions T$^3$RL as a verified online data synthesizer, highlighting the importance of tool verification for reliable online adaptation and demonstrating extensibility to other verifiable domains.
By Ruotong Liao, Nikolai R\"ohrich, Xiaohan Wang, Yuhui Zhang, Yasaman Samadzadeh, Volker Tresp, Serena Yeung-Levy
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
Latent-CoT-Drive (LCDrive) is a vision‑language‑action model for autonomous driving that replaces natural‑language chain‑of‑thought reasoning with a latent language capturing possible outcomes of driving actions. The model interleaves action‑proposal tokens, aligned with the model’s output actions, and world‑model tokens grounded in a learned latent world model to reason about future outcomes. After a supervised cold‑start using ground‑truth future rollouts, LCDrive is further refined with closed‑loop reinforcement learning, achieving faster inference, higher‑quality trajectories, and greater gains from interactive RL than both non‑reasoning and text‑reasoning baselines on a large‑scale end‑to‑end driving benchmark.
By Shuhan Tan, Kashyap Chitta, Yuxiao Chen, Ran Tian, Yurong You, Yan Wang, Wenjie Luo, Yulong Cao, Philipp Krahenbuhl, Marco Pavone, Boris Ivanovic
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
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
The paper investigates how different estimators of the reverse Kullback–Leibler (KL) divergence used as a regularization term in reinforcement learning (RL) training of large language models (LLMs) affect training stability and downstream performance. By analyzing gradient bias across various estimator configurations, the authors demonstrate that biased gradients can cause training instabilities, while unbiased configurations improve performance on both in‑domain and out‑of‑domain tasks. Experiments on Qwen2.5‑7B, Llama‑3.1‑8B‑Instruct, and Qwen3‑4B‑Instruct‑2507 confirm these findings and show that KL regularization also stabilizes off‑policy RL training in asynchronous setups.
By Vedant Shah, Johan Obando-Ceron, Vineet Jain, Brian Bartoldson, Bhavya Kailkhura, Sarthak Mittal, Glen Berseth, Pablo Samuel Castro, Yoshua Bengio, Esmeralda S. Whitammer, Moksh Jain, Siddarth Venkatraman, Aaron Courville
The paper introduces PLUS, a framework that uses reinforcement learning to generate text-based summaries of individual users’ preferences, characteristics, and past conversations. These summaries condition a reward model, allowing it to predict personalized response preferences and improving reward accuracy by 11–77 % over the standard Bradley‑Terry model. PLUS demonstrates robust performance with new users and topics, achieves a 25 % improvement over existing personalized RLHF techniques, and enables zero‑shot personalization for state‑of‑the‑art models like GPT‑4.
By Hyunji Nam, Yanming Wan, Mickel Liu, Peter Ahnn, Jianxun Lian, Natasha Jaques
ClueWeaver is a dual-agent framework designed to enable compact, locally deployable language models to answer questions about long literary narratives. The Finder agent retrieves passages that contain answer-critical clues, while the Interpreter agent derives the answer from those passages, generates rationales with paragraph-ID citations, and performs self-calibration for high-risk questions. Both agents are trained with reward-guided reinforcement learning to prioritize evidence retention, correctness, grounding, and concise explanations, resulting in improved performance and inspectability over end-to-end prompting.
By Jihao Zhu, Zhiwei Yang, Wenxiao Zhang, Junqian Zhao, Qi You, Fangqi Wang, Zheyuan Deng, Hanzhe Yang, Yu Liu, Jin B. Hong
InternBootcamp is an open‑source framework that offers over 1,000 domain‑diverse task environments for large language model (LLM) reasoning research. It introduces Bootcamp‑Eval, an automatically generated benchmark for comprehensive performance assessment. Experiments show that training on InternBootcamp significantly improves reasoning performance, with a 32B model achieving state‑of‑the‑art results on Bootcamp‑Eval and other established benchmarks, demonstrating that scaling the number of training tasks yields consistent gains.
By Peiji Li, Jiasheng Ye, Yongkang Chen, Linyang Li, Yichuan Ma, Zijie Yu, Ganqu Cui, Haozhan Li, Jiacheng Chen, Chengqi Lyu, Wenwei Zhang, Qipeng Guo, Dahua Lin, Bowen Zhou, Kai Chen
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 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 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 an agent‑based modeling framework that integrates a physical infrastructure layer, a V2X connectivity layer, and a decision layer using reinforcement learning and multi‑agent reinforcement learning to simulate smart freight corridors. Three scenarios—Baseline, Assisted, and Cognitive—are evaluated on throughput, congestion, energy, emissions, and robustness, with the Cognitive scenario outperforming the baseline in throughput and congestion, and the Assisted scenario achieving energy savings via platooning. Sensitivity analysis shows that the smart corridor’s throughput advantage grows under high demand and that MARL coordination better utilizes fixed charging capacity than rule‑based methods.
By Madelaine Martinez-Ferguson, Chun Wang, Mustafa Can Camur, Xueping Li