MAGMA-GEN is an on‑policy data‑generation pipeline that transforms ambiguous failures in hierarchical robotic manipulation into validated recovery supervision. It uses a privileged coach to hypothesize early decision‑level errors and proposes localized corrections, then retains only those candidates that improve downstream progress when re‑executed from the same state. This approach generates supervised examples from the agent’s own failure distribution, enabling improved task success and recovery without requiring per‑step human demonstrations.
By Loan Bernat (LAAS-GEPETTO), Matthieu Grard (LAAS-RAP), Ariane Herbulot (LAAS-RAP), Florent Lamiraux (LAAS-GEPETTO)
DeliveryGym is a 3D reinforcement learning environment that simulates continuous courier shifts, integrating multimodal tool interaction, persistent world dynamics, and trajectory‑based rewards derived from simulator events. It allows agents to learn how their decisions affect time, energy, and money across an entire shift, and it adapts future training shifts to the policy’s weaknesses while keeping evaluation fixed. Experiments on six models and 13 city maps show a significant gap between task execution and optimal sequencing, with RL improving Qwen3‑VL‑4B’s net income by 54.3% and adaptive training boosting test income by 16.5% over uniform sampling.
By Haoqiang Kang, Yiming Zhang, Yiyang Guo, Chuying Li, Jianzhi Shen, Tianruo Rose Xu, Xiaokang Ye, Lianhui Qin
MATCH is a closed‑loop framework for model‑aware tool learning that combines curriculum scheduling with hierarchically gated rewards. It introduces Model‑Aware Curriculum Learning (MACL), which dynamically adjusts sample difficulty based on reward signals, and Hierarchical Tool‑call Gated Reward (HTGR), which allocates credit at the tool name, argument key, and argument value levels only when prerequisites are met. Experiments on API‑Bank and BFCL V3 show MATCH achieving 72.19% and 62.87% overall accuracy, outperforming both supervised and RL‑based baselines across multiple backbone models.
By Shihao Liu, Hao Yin, Lijun Liu, Zhengzong Chen, Yuanyuan Zhao, Fei Huang
The paper introduces VLN on the Fly, an onboard vision‑language navigation stack for aerial robots that separates grounding, planning, and control into inspectable stages. A quantized vision‑language model grounds instructions to a coarse image cell, depth estimation lifts this to a 3D goal, a fast B‑spline planner generates a feasible trajectory, and a pretrained reinforcement learning policy translates the trajectory into motor commands. In controlled indoor flights, the stack achieved the target in 13 of 15 trials with a mean goal error of 5.72 cm and 39.3% GPU utilization, and successfully tracked collision‑free trajectories in cluttered environments.
By Marco S. Tayar, Felipe Tommaselli, Gianluca Capezutto, Pedro Antonio Rabelo Saraiva, Pedro H. V. de Freitas, Lucas Kido, Guilherme Sonego, Ricardo V. Godoy, Marcelo Becker
RetireOPD introduces a self-retiring on‑policy distillation method for agentic reinforcement learning. It first trains a skill‑conditioned teacher with environment rewards, then jointly trains a skill‑free student with RL and OPD, allowing the student to autonomously stop using the teacher when its performance aligns with the teacher’s. Experiments on Qwen2.5 models show significant gains in ALFWorld success rates and WebShop accuracy compared to RL baselines and the teacher itself.
By Yan Yu, Zhengxi Lu, Yizhou Liu, Yichen Pan, Aozhe Wang, Qipeng Chen, Hua Yang, Wenqi Zhang, Weiming Lu, Qianglong Chen, Yongliang Shen
TripScore is a benchmark and evaluation framework for large language models (LLMs) in travel planning, built from real user logs and calibrated with 1,468 pairwise judgments from 203 travel experts. It uses a hierarchical feasibility gate for format and commonsense checks, and a unified point-wise reward that combines soft quality and preference fulfillment. Experiments show that reinforcement learning fine‑tuning, such as GRPO, consistently outperforms other methods when evaluated with TripScore.
By Yincen Qu, Huan Xiao, Feng Li, Gregory Li, Hui Zhou, Xiangying Dai, Xiaoru Dai, Xuan Huang
The paper introduces Reward Stimulation Implicit Q-Learning (RSIQL), a non-hierarchical approach to improve offline goal-conditioned reinforcement learning. RSIQL adds auxiliary reward signals at intermediate states that are predicted to aid progress toward the goal, thereby reducing the delay in training supervision. Experiments on D4RL goal-reaching benchmarks and OGBench demonstrate that RSIQL outperforms baseline goal-conditioned IQL and rivals hierarchical offline methods while maintaining a simple flat policy structure.
By Jing Zhang
The paper introduces Robust Fed-Q, a federated Q‑learning algorithm designed for settings where multiple agents interact with a shared Markov Decision Process and communicate through a central server. It combines model‑based and model‑free reinforcement learning techniques with a median‑of‑means strategy from robust statistics to handle a small fraction of adversarial agents. The authors prove that Robust Fed-Q achieves exact convergence to the optimal value function with high probability, attains near‑optimal finite‑time rates that benefit from collaboration, and requires only “~O(1)” communication rounds per guarantee.
By Sreejeet Maity, Aritra Mitra
The paper introduces BADA, a Boundary-Aware Data Augmentation technique for offline reinforcement learning. By interpolating neighboring states to create synthetic data that respects the original distribution, BADA improves in-distribution generalization and robustness. Experiments on limited offline datasets show that BADA achieves state-of-the-art performance across diverse benchmarks.
By Gong Gao, Weidong Zhao, Xianhui Liu
The paper introduces a method called score centering to address the training‑inference mismatch (TIM) that destabilizes reinforcement learning for large language models. By adding an additive correction term that cancels drift between training and inference engines, score centering stabilizes RL and can match or surpass importance‑sampling techniques, especially as model size and mismatch severity increase. The approach also composes with importance sampling, yielding further performance gains in staleness experiments.
By Martin Marek, Max Ryabinin
EmbodiedMind introduces a three-stage training paradigm for embodied foundation models that tackles inefficient sample use, task imbalance, and credit assignment in long-horizon planning. The stages—Rejection Sampling-based Fine‑Tuning, Iterative Rejection GRPO, and Trie‑GRPO—filter low‑informative data, balance task difficulty, and use action prefix trees for step‑level advantage estimation. This approach yields a state‑of‑the‑art average performance of 70.02% across 18 benchmarks, notably improving long‑horizon task planning accuracy.
By Feifan Wang, Zongbing Zhang, Yu Zhang, Lingfeng Wang, Yurui Zhu, Jin Deng, Mingliang Zhang, Zhengguang Gao, Yongcheng Wang, Jin Xu, Ri Yang
MILER is an end‑to‑end reinforcement learning framework that achieves zero‑shot sim‑to‑real transfer for autonomous driving in unstructured environments. It uses a custom semantic mid‑level representation (MLR) simulator for offline training, and during deployment it processes real camera and LiDAR data with BEVFusion to produce a compatible bird’s‑eye‑view representation. The policy’s actions are applied via a trajectory‑alignment strategy, allowing the system to drive 17.3 km on a 3.0 km test track without human intervention, all running on a Jetson AGX Orin.
By Thomas Steinecker, Denis Trescher, Alexander Bienemann, Thorsten Luettel, Mirko Maehlisch
OPTED is a method for on‑policy fine‑tuning of end‑to‑end driving models that separates reinforcement learning from the policy update. A privileged teacher trained with RL on vectorized inputs (HD‑maps and bounding boxes) supervises the pre‑trained student during closed‑loop post‑training. Applied to the camera‑based models TransFuser and VaVAM in AlpaSim, OPTED boosts driving scores by 1.6× and 9.5×, respectively, while requiring roughly three orders of magnitude fewer simulator interactions than direct RL post‑training.
By Damiano Da Col, Maximilian Igl, Peter Karkus, Kashyap Chitta, Boris Ivanovic, Marco Pavone, Konrad Schindler, Christos Sakaridis
The paper introduces QUATRO, a reinforcement‑learning approach for fine‑tuning large language models that enforces trust‑region constraints directly rather than relying on heuristic clipping. By deriving a principled objective, QUATRO provides explicit control over policy updates and stabilizes entropy during training. Experiments on mathematical reasoning benchmarks demonstrate that QUATRO maintains stable training even with higher learning rates and increased policy staleness.
By Doyeon Lee, Eunyi Lyou, Hyunsoo Cho, Sookyung Kim, Joonseok Lee, Jaemoo Choi
The paper investigates how alignment training, specifically reinforcement learning from human feedback (RLHF), affects the internal partisan structure of a large language model. Using a mechanistic case study on Llama 3.1 8B, the authors find that alignment training does not erase the model’s partisan geometry but compresses its variance, producing consistently balanced, non‑partisan outputs. Sparse autoencoder analysis and feature‑level steering experiments reveal that policy‑encoding features become inactive in the aligned model, indicating a causal disconnect rather than structural removal of partisan knowledge.
By Wendy K. Tam
AnyviewMeter is a geometry‑conditioned adaptation framework that fine‑tunes pretrained robotic reward models to account for camera viewpoint and occlusion. It uses low‑rank fine‑tuning, token‑aligned Plücker rays, and synchronous block attention to fuse synchronized multi‑view observations. Experiments on PickCube and simulated manipulation tasks show significant reductions in mean absolute error and improved temporal ordering compared to single‑view or RGB‑only fine‑tuning.
By Yuang Tu, Runjia Tan, Yujie Yan, Jinghan Hu, Chen Lv
The paper introduces AttnSign, a vision‑language model that improves sign language translation by steering spatial‑temporal attention. It first supervises attention on sign‑relevant regions such as faces and hands in each frame, then uses an RL‑based motion‑cadence method to focus on keyframes. Experiments on How2Sign and OpenASL show AttnSign consistently outperforms existing methods.
By Meibo Hu, Guohao Sun, Annemarie D. Ross, Sheng Li, Zhiqiang Tao
The paper presents a portfolio-based approach to constrained multi-objective Bayesian optimization for materials design, framing acquisition‑function selection as an adaptive policy problem. Two controllers—UCB‑Bandit, a modified UCB multi‑armed bandit, and Agentic‑Switch, a multi‑agent system powered by a large language model—were tested against fixed‑policy baselines on synthetic benchmarks and two real materials design case studies. The adaptive policies achieved competitive results in cumulative feasibility counts and feasible hypervolume improvement, outperforming individual acquisition functions that excelled only in a single metric.
By Sushant Sinha, Christofer Hardcastle, Robert Robinson, Shakti Prasad Padhy, Brent Vela, Douglas Allaire, Raymundo Arroyave
The paper introduces a hierarchical hybrid architecture combining large language models (LLMs) and multi-agent reinforcement learning (MARL) to manage heterogeneous unmanned aerial systems in low‑altitude wireless networks (LAWNs). An outer LLM‑driven loop interprets service requirements and operator intent to reconfigure objectives and resource priorities, while an inner MARL loop executes decentralized policies under the updated game. A logistics‑monitoring case study demonstrates the framework’s ability to coordinate diverse services and adapt to changing conditions without retraining the MARL policies.
By Nguyen Duc Minh Quang, Chang Liu, Shuangyang Li, Derrick Wing Kwan Ng
The paper introduces a new evaluation protocol called checkpoint handoff to disentangle the contributions of reaching a target state and solving the task in reinforcement learning agents. By cloning states reached by one checkpoint and handing them to another without retraining, the authors separate the REACH metric (how often a policy arrives at a state confirmed to be a fixed number of actions from success) from the SOLVE metric (how often it finishes from that identical state). Across two benchmarks and pipelines, the analysis shows that RL history benefits RL solvers more than SFT solvers, and that independent REACH and SOLVE gaps predict overall performance.
By Xuan Liu, Jingbin Qian