MARS: Reinforcing Multi-Agent Reasoning of LLMs through Self-Play in Strategic Games

• URL: http://arxiv.org/abs/2510.15414v1
• Date: Fri, 17 Oct 2025 08:08:06 GMT
• Title: MARS: Reinforcing Multi-Agent Reasoning of LLMs through Self-Play in Strategic Games
• Authors: Huining Yuan, Zelai Xu, Zheyue Tan, Xiangmin Yi, Mo Guang, Kaiwen Long, Haojia Hui, Boxun Li, Xinlei Chen, Bo Zhao, Xiao-Ping Zhang, Chao Yu, Yu Wang,
• Abstract summary: We introduce MARS, an end-to-end RL framework that incentivizes Multi-Agent Reasoning of LLMs through Self-play in both cooperative and competitive games.<n>Our results establish end-to-end RL training with self-play in strategic games as a powerful approach for developing generalizable multi-agent reasoning capabilities in LLMs.
• Score: 30.876486250077956
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Developing Large Language Models (LLMs) to cooperate and compete effectively within multi-agent systems is a critical step towards more advanced intelligence. While reinforcement learning (RL) has proven effective for enhancing reasoning in single-agent tasks, its extension to multi-turn, multi-agent scenarios remains underexplored due to the challenges of long-horizon credit assignment and agent-specific advantage estimation. To address these challenges, we introduce MARS, an end-to-end RL framework that incentivizes Multi-Agent Reasoning of LLMs through Self-play in both cooperative and competitive games. MARS features a turn-level advantage estimator that aligns learning signals with each interaction for credit assignment, and an agent-specific advantage normalization to stabilize multi-agent training. By learning with self-play across cooperative and competitive games, the MARS agent trained from Qwen3-4B develops strong strategic abilities that generalize to held-out games with up to 28.7% performance improvements. More importantly, the capability acquired through self-play generalizes beyond games, yielding consistent performance gains of multi-agent systems in reasoning benchmarks. When integrated into leading multi-agent systems, our MARS agent achieves significant performance gains of 10.0% on AIME and 12.5% on GPQA-Diamond. These results establish end-to-end RL training with self-play in strategic games as a powerful approach for developing generalizable multi-agent reasoning capabilities in LLMs. Our code and models are publicly available at https://github.com/thu-nics/MARS.

Related papers

• MARTI-MARS$^2$: Scaling Multi-Agent Self-Search via Reinforcement Learning for Code Generation [64.2621682259008]
  Multi-Agent Reinforced Training and Inference Framework with Self-Search Scaling (MARTI-MARS2)<n>We propose a Multi-Agent Reinforced Training and Inference Framework with Self-Search Scaling (MARTI-MARS2) to integrate policy learning with multi-agent tree search.<n>We show that MARTI-MARS2 achieves 77.7%, outperforming strong baselines like GPT-5.1 on challenging code generation benchmarks.
  arXiv  Detail & Related papers  (2026-02-08T07:28:44Z)
• Collaborative Multi-Agent Test-Time Reinforcement Learning for Reasoning [112.16686518063456]
  We introduce textbfMulti-Agent Test-Time Reinforcement Learning (MATTRL), a framework that injects structured textual experience into multi-agent deliberation at inference time.<n>MATTRL forms a multi-expert team of specialists for multi-turn discussions, retrieves and integrates test-time experiences, and reaches consensus for final decision-making.<n>Across challenging benchmarks in medicine, math, and education, MATTRL improves accuracy by an average of 3.67% over a multi-agent baseline, and by 8.67% over comparable single-agent baselines
  arXiv  Detail & Related papers  (2026-01-14T17:57:43Z)
• Learning Robust Social Strategies with Large Language Models [7.697496386429445]
  Reinforcement learning is effective for aligning large language models (LLMs) in the single-agent regime.<n>We show that standard RL in multi-agent settings often converges to defecting, self-interested policies.<n>To address this tendency of RL to converge to poor equilibria, we adapt a recent opponent-learning awareness algorithm, Advantage Alignment.
  arXiv  Detail & Related papers  (2025-11-24T18:43:46Z)
• AgentGym-RL: Training LLM Agents for Long-Horizon Decision Making through Multi-Turn Reinforcement Learning [129.44038804430542]
  We introduce AgentGym-RL, a new framework to train LLM agents for multi-turn interactive decision-making through RL.<n>We propose ScalingInter-RL, a training approach designed for exploration-exploitation balance and stable RL optimization.<n>Our agents match or surpass commercial models on 27 tasks across diverse environments.
  arXiv  Detail & Related papers  (2025-09-10T16:46:11Z)
• How to Train a Leader: Hierarchical Reasoning in Multi-Agent LLMs [16.853362180877593]
  We introduce a hierarchical multi-agent framework that trains only a single leader LLM to coordinate a team of untrained peer agents.<n>Our results highlight the effectiveness and efficiency of training a single, flexible leader for collaborative reasoning in multi-agent LLM systems.
  arXiv  Detail & Related papers  (2025-07-11T18:34:07Z)
• ReMA: Learning to Meta-think for LLMs with Multi-Agent Reinforcement Learning [53.817538122688944]
  We introduce Reinforced Meta-thinking Agents (ReMA) to elicit meta-thinking behaviors from Reasoning of Large Language Models (LLMs)<n>ReMA decouples the reasoning process into two hierarchical agents: a high-level meta-thinking agent responsible for generating strategic oversight and plans, and a low-level reasoning agent for detailed executions.<n> Empirical results from single-turn experiments demonstrate that ReMA outperforms single-agent RL baselines on complex reasoning tasks.
  arXiv  Detail & Related papers  (2025-03-12T16:05:31Z)
• MALT: Improving Reasoning with Multi-Agent LLM Training [67.76186488361685]
  MALT (Multi-Agent LLM Training) is a novel post-training strategy that divides the reasoning process into generation, verification, and refinement steps.<n>On MATH, GSM8K, and CSQA, MALT surpasses the same baseline LLM with a relative improvement of 15.66%, 7.42%, and 9.40% respectively.
  arXiv  Detail & Related papers  (2024-12-02T19:30:36Z)
• Learning From Good Trajectories in Offline Multi-Agent Reinforcement Learning [98.07495732562654]
  offline multi-agent reinforcement learning (MARL) aims to learn effective multi-agent policies from pre-collected datasets. One agent learned by offline MARL often inherits this random policy, jeopardizing the performance of the entire team. We propose a novel framework called Shared Individual Trajectories (SIT) to address this problem.
  arXiv  Detail & Related papers  (2022-11-28T18:11:26Z)
• Group-Agent Reinforcement Learning [12.915860504511523]
  It can largely benefit the reinforcement learning process of each agent if multiple geographically distributed agents perform their separate RL tasks cooperatively. We propose a distributed RL framework called DDAL (Decentralised Distributed Asynchronous Learning) designed for group-agent reinforcement learning (GARL)
  arXiv  Detail & Related papers  (2022-02-10T16:40:59Z)
• Strategically Efficient Exploration in Competitive Multi-agent Reinforcement Learning [25.041622707261897]
  This work seeks to understand the role of optimistic exploration in non-cooperative multi-agent settings. We will show that, in zero-sum games, optimistic exploration can cause the learner to waste time sampling parts of the state space that are irrelevant to strategic play. To address this issue, we introduce a formal notion of strategically efficient exploration in Markov games, and use this to develop two strategically efficient learning algorithms for finite Markov games.
  arXiv  Detail & Related papers  (2021-07-30T15:22:59Z)
• Learning to Incentivize Other Learning Agents [73.03133692589532]
  We show how to equip RL agents with the ability to give rewards directly to other agents, using a learned incentive function. Such agents significantly outperform standard RL and opponent-shaping agents in challenging general-sum Markov games. Our work points toward more opportunities and challenges along the path to ensure the common good in a multi-agent future.
  arXiv  Detail & Related papers  (2020-06-10T20:12:38Z)
• Scalable Multi-Agent Inverse Reinforcement Learning via Actor-Attention-Critic [54.2180984002807]
  Multi-agent adversarial inverse reinforcement learning (MA-AIRL) is a recent approach that applies single-agent AIRL to multi-agent problems. We propose a multi-agent inverse RL algorithm that is more sample-efficient and scalable than previous works.
  arXiv  Detail & Related papers  (2020-02-24T20:30:45Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.