Dr. MAS: Stable Reinforcement Learning for Multi-Agent LLM Systems

TL;DR

This paper introduces Dr. MAS, a stable reinforcement learning method for multi-agent LLM systems that normalizes advantages per agent, leading to more reliable training and improved performance on reasoning and search benchmarks.

Contribution

The paper identifies a key instability in multi-agent RL training and proposes Dr. MAS, a normalization-based approach that stabilizes training and enhances multi-agent LLM system performance.

Findings

Dr. MAS significantly improves performance on math reasoning and search benchmarks.

It reduces gradient spikes and stabilizes training in multi-agent LLM systems.

Effective under heterogeneous agent-model configurations.

Abstract

Multi-agent LLM systems enable advanced reasoning and tool use via role specialization, yet reliable reinforcement learning (RL) post-training for such systems remains difficult. In this work, we theoretically pinpoint a key reason for training instability when extending group-based RL to multi-agent LLM systems. We show that under GRPO-style optimization, a global normalization baseline may deviate from diverse agents' reward distributions, which ultimately leads to gradient-norm instability. Based on this finding, we propose Dr. MAS, a simple and stable RL training recipe for multi-agent LLM systems. Dr. MAS uses an agent-wise remedy: normalizing advantages per agent using each agent's own reward statistics, which calibrates gradient scales and dramatically stabilizes training, both theoretically and empirically. Beyond the algorithm, Dr. MAS provides an end-to-end RL training framework for multi-agent LLM systems, supporting scalable orchestration, flexible per-agent LLM serving and optimization configs, and shared resource scheduling of LLM actor backends. We evaluate Dr. MAS on multi-agent math reasoning and multi-turn search benchmarks using Qwen2.5 and Qwen3 series models. Dr. MAS achieves clear gains over vanilla GRPO (e.g., +5.6\% avg@16 and +4.6\% pass@16 on math, and +15.2\% avg@16 and +13.1\% pass@16 on search) while largely eliminating gradient spikes. Moreover, it remains highly effective under heterogeneous agent-model assignments while improving efficiency.