The paper introduces LUGL (Local Updates, Global Learning), a framework that separates data collection from model fitting, allowing non‑incremental learners such as gradient‑boosted trees (LightGBM) to be used in reinforcement learning for games. LUGL alternates between a local update phase—where agents play self‑play games and store tabular updates—and a global learning phase—where a function approximator is trained on the accumulated table before it is reset. Experiments on both perfect‑information and imperfect‑information games show that LightGBM‑based agents perform competitively or better than neural‑network baselines like DQN and DeepCFR.
By David Milec, Spyridon Samothrakis, Michael Fairbank, Dennis J. N. J. Soemers
The paper reviews how Large Language Models (LLMs) are being adapted for medical reasoning, moving beyond single-step answers to systems that can systematically, transparently, and verifiably reason. It introduces a taxonomy of enhancement techniques, split into training-time methods such as supervised fine‑tuning and reinforcement learning, and test-time methods like prompt engineering and multi‑agent systems. The review examines their application across text, image, and code modalities in key clinical areas—diagnosis, education, and treatment planning—and tracks the shift in evaluation benchmarks from simple accuracy to more nuanced assessments of reasoning quality and visual interpretability.
By Zizhan Ma, Wenxuan Wang, Meidan Ding, Shiyi Zheng, Shengyuan Liu, Jie Liu, Jiaming Ji, Linlin Shen, Yixuan Yuan, Wenting Chen
The paper explores a runtime strategy-selection framework where a large language model (LLM) guides a pre‑trained reinforcement learning (RL) policy for non‑player characters (NPCs) in a Unity combat game without altering the underlying policy. Five NPC agents sharing a PPO policy were compared in a baseline setup and an LLM‑augmented setup, where a locally hosted Mistral 7B model assigns one of four tactical tags every five seconds based on live game state. Across 600 episodes against three scripted opponents, the LLM‑augmented agents more than doubled their win rate against a Balanced opponent, improved performance against an Evasive opponent, but struggled against an Aggressive opponent due to over‑reliance on encirclement; analysis of 2,430 strategy selections revealed limited zero‑shot differentiation with the model favoring Surround in 83.8% of cases.
By Hrithika Deepu Nair, Kayvan Karim
RL-ADA introduces a co‑evolutionary training framework that replaces costly human annotations with world‑feedback rewards derived from interaction outcomes. In this system, a large Customer Support Agent and an Adversarial Customer Agent train together, guided by an automated judge that rewards successful resolution and realistic intent‑concealing utterances, respectively. Applied to a banking support proof of concept, the method eliminates routing errors and doubles the end‑to‑end PASS rate over five cycles, while also revealing a new adversarial strategy called Contextual Camouflage.
By Ram Narayanan, Harshit Rajgarhia, Abhishek Mukherji
ToPO (Token-Oriented Preference Optimization) is a method for attention-based latent diffusion models that constructs a per‑minibatch, detached, separable spatial‑temporal route using branchwise squared‑residual contrast from a frozen reference denoiser. It employs preferred‑branch cross‑attention with content tokens to modulate spatial factors and adds an auxiliary pixel‑midpoint ordering term, all without local labels or a learned reward model. In experiments with shared update schedules, ToPO outperforms Diffusion‑DPO on five SD‑1.5 metrics and on HPSv2, ImageReward, and CLIP for SDXL, and achieves higher raw win shares in a blind SDXL A/B study.
By Juntao Xu, Shihong Li, Hoi Fan Au, Ning Zhu
StrixAE is an intelligent audio enhancement agent that uses a multimodal large language model to coordinate multiple enhancement and personalization models. It is trained in two stages: first, chain‑of‑thought supervised fine‑tuning on AcoustBench, and second, Audio Perception Reinforcement Learning that optimizes format validity, structural coherence, and perceptual quality. The approach yields state‑of‑the‑art performance and strong generalization on real‑world test datasets, outperforming many existing open‑source and proprietary solutions.
By Chenglin Wu, Junjie Wu, Jinhang Chen, Mingyang Chen, Zixu Lin, Jiabian Chen, Xinghao Ding, Xiaotong Tu
The paper investigates zero‑shot task generalisation in offline multi‑agent reinforcement learning by extending sequence‑modeling architectures to support multi‑task observation and action spaces and variable agent counts. It finds that increasing task diversity, rather than merely enlarging the dataset, is the key driver for robust zero‑shot transfer. Experiments on four challenging environments show a 3.2× mean improvement on held‑out tasks compared to single‑task models and outperform strong behaviour‑cloning baselines.
By Oussama Hidaoui, Omer Ebead, Ulrich Armel Mbou Sob, Siddarth Singh, Juan Claude Formanek, Felix Chalumeau, Omayma Mahjoub, Sasha Abramowitz, Ruan John de Kock, Wiem Khlifi, Louay Ben Nessir, Simon Verster Du Toit, Daniel Rajaonarivonivelomanantsoa, Asim Awad Osman, Arnol Manuel Fokam, Refiloe Shabe, Arnu Pretorius
PaperScout is an autonomous agent that treats academic paper search as a sequential decision-making process, dynamically deciding when and how to use search and expansion tools based on accumulated context. The authors identify a granularity mismatch in standard reinforcement learning for multi-turn tasks and propose Proximal Sequence Policy Optimization (PSPO), a sequence-level policy optimization method that aligns learning with agent–environment interactions. Experiments on synthetic and real-world benchmarks show that PaperScout outperforms structured retrieval and RL baselines in recall and relevance, demonstrating the effectiveness of its adaptive agentic framework and optimization strategy.
By Tingyue Pan, Jie Ouyang, Mingyue Cheng, Qingchuan Li, Zirui Liu, Daoyu Wang, Mingfan Pan, Shuo Yu, Qi Liu, Enhong Chen
The paper introduces F-GRPO, a method that addresses the issue of reinforcement learning policies overfitting to common trajectories while neglecting rare correct ones. By deriving the probability of prompt‑local tail‑miss events and proposing a difficulty‑aware scaling coefficient inspired by Focal loss, the authors show that down‑weighting high‑success sampled groups can improve performance. Experiments on categorical simulations, Maze, and large language models (Qwen2.5‑7B) demonstrate that F‑GRPO raises average math pass rates and out‑of‑distribution performance without increasing group size or computational cost.
By Daniil Plyusov, Alexey Gorbatovski, Boris Shaposhnikov, Viacheslav Sinii, Alexey Malakhov, Daria Korotyshova, Daniil Gavrilov
The paper introduces Test Cases Scaling (TCS), a two‑stage reinforcement learning framework that automatically generates sound and discriminative test cases for code large language models (LLMs). Stage 1 trains a test generator to produce tests aligned with a reference solution, while Stage 2 refines the generator to produce counterexample tests that target the solver’s current failure modes. Experiments on TACO and LiveCodeBench show that TCS improves pass@1 and inference‑time answer selection, and the learned test generator can also effectively rank outputs from other LLMs.
By Jiacheng Xu, Wentao Zhang, Zhiyi Lyu, Fuxiang Zhang, Chaojie Wang, Yang Liu, Bo An
The paper introduces FactoSR, a factorized reinforcement learning framework designed to improve spatial reasoning in Vision‑Language Models by addressing a dimensional mismatch between 2D visual inputs and the 3D+temporal nature of the physical world. FactoSR decomposes the reasoning task into three orthogonal geometric sub‑objectives—planar correspondence (XY), depth consistency (Z), and temporal reversibility (T)—and optimizes these constraints within a unified policy learning mechanism. Experiments on multi‑view and video benchmarks show that this decomposition yields significant performance gains, achieving a 5.9% improvement on VSI‑Bench and 4.5% on All‑Angles‑Bench.
By Yijun Yang, Shenghe Zheng, Wenbo Li, Jianhui Liu, Haoze Sun, Yanbing Zhang, Jiaxiu Jiang, Lin Song, Haoyang Huang, Nan Duan, Lei Zhu
The paper introduces MINT‑Safe, a new open‑source dataset of 11,270 multi‑image dialogues and 500 refusal VQA pairs designed to expose safety risks in multi‑modal large language models during open‑ended conversations. It also proposes TAD‑Align, a turn‑aware dual‑objective reward framework that dynamically up‑weights dialogue turns with inconsistent safety behavior, improving safety metrics on Qwen2.5‑VL‑7B‑Instruct and LLaVA‑Next‑7B. The results show over 10% reduction in attack success rate and notable gains in harmlessness and helpfulness while maintaining overall model performance.
By Han Zhu, Jiale Chen, Chengkun Cai, Shengjie Sun, Haoran Li, Yujin Zhou, Chi-Min Chan, Pengcheng Wen, Lei Li, Yike Guo, Sirui Han
The paper introduces "environment evolution," a method that incrementally raises the difficulty of interactive environments off‑policy, scheduling their generation across training generations to supply continuous learning signals. It derives three evolution directions tied to a multi‑turn learning objective and implements them via a loop‑engineered multi‑agent harness. Experiments with models such as Hy4 preview, Claude Opus 5, GPT‑5.6 Sol, Qwen3.6‑27B, and Qwen3.6‑35B‑A3B demonstrate that this approach consistently creates harder environments and boosts terminal‑agent performance on Terminal‑Bench 2.1 by 14.4–18.0 percentage points.
By Zhiyuan Fan, Tinghao Yu, Yuanjun Cai, Jiang Zhou, Jiangtao Guan, Jincheng Liu, Yun Yang, Dingxin Hu, Zhuo Han, Xing Wu, Feng Zhang, Lilin Wang
The paper presents an end‑to‑end framework that uses constraint‑oriented hypergraphs and reinforcement learning to solve vehicle routing problems. It introduces a dynamic hyperedge reconstruction strategy for better hypergraph representation and a double‑pointer attention decoder for iterative solution generation. Experiments on benchmark datasets show that the method removes the need for complex heuristic operators while improving solution quality.
By Zhenwei Wang, Tiehua Zhang, Jing Liu, Heng Yu, Kaizhu Huang, Ruibin Bai
The paper offers a new way to view the occupancy measure in reinforcement learning by embedding the planning criterion into the dynamics via a resetting planning process. The resulting stationary measure, called the visitation measure, forms a dually flat statistical manifold with two affine charts: visitation probabilities and log-policies, which are dual under conditional entropy. This geometric framework allows planning-as-inference to extend beyond linear rewards to nonlinear functionals of visitation, with each iteration solvable by a natural-gradient step and provides a new interpretation of the temporal-difference error as a marginal-utility estimate.
By Nikola Milosevic, Asaki Kataoka, Nicolas Hinrichs, Kenji Doya, Nico Scherf
The paper introduces a formal framework that uses symmetries in data to keep causal mechanisms invariant, providing a simple and general mathematical language for causal reasoning. It outlines how to describe models and queries, and presents strategies for rigorously identifying causal effects from data within this framework. The approach reproduces known results for IID data and extends causal analysis to non‑IID settings, complex queries beyond do‑ or soft‑interventions, and incorporates missing data, transfer, and robustness considerations.
By Martin Rabel, Jakob Runge
The paper introduces "environment evolution," a method that incrementally raises the difficulty of interactive environments for terminal agents using off‑policy updates. It derives three evolution directions tied to a multi‑turn learning objective and implements them via a loop‑engineered multi‑agent harness. Experiments with models such as Hy4 preview, Claude Opus 5, GPT‑5.6 Sol, Qwen3.6‑27B, and Qwen3.6‑35B‑A3B demonstrate that this approach consistently generates harder environments and yields significant performance gains on Terminal‑Bench 2.1.
The paper investigates how to combine on‑policy distillation (OPD) and reinforcement learning with verifiable rewards (RLVR) for post‑training reasoning in large language models. It finds that a simple two‑stage approach—first applying OPD, then switching to RL—outperforms single‑signal methods and other joint baselines on logic and math reasoning benchmarks. The authors explain this advantage by showing that OPD broadens the model’s coverage of teacher‑supported solutions while RL sharpens performance within that support, and they provide practical guidance on when to switch based on OPD validation scores.
DRACO introduces a method for fine‑grained credit assignment in long‑horizon reinforcement learning tasks that lack verifiable rewards. It dynamically generates multi‑criteria rubrics during training, scores them once per trajectory, and redistributes the resulting judgment over the steps responsible for each rubric to produce differentiated per‑step advantages. Experiments on AppWorld and Tau‑Bench show that DRACO outperforms baseline models and other rubric‑based approaches, achieving significant performance gains without relying on verifiers.
The paper identifies a flaw in Group Relative Policy Optimization (GRPO) called the spurious advantage, where the advantage estimator can incorrectly reward rollouts that reach correct answers by guessing rather than reasoning. This issue appears in bounded-answer tasks, open-answer sets with bounded sub-cases, and search agents with many paths to the same answer. The authors propose SIGNBALANCE, a new magnitude estimator that removes composition, maintains verifier sign, applies a global scale, and uses stop‑gradient per‑class rescaling to restore zero‑mean balance, achieving comparable or better performance on math and search benchmarks.