Dr. MAS: Stable Reinforcement Learning for Multi-Agent LLM Systems
This addresses the problem of reliable RL training for multi-agent LLM systems, which is incremental as it builds on existing GRPO-style optimization with a specific stabilization improvement.
The paper tackles training instability in multi-agent LLM systems during reinforcement learning post-training by identifying that global normalization baselines cause gradient-norm instability due to diverse agent reward distributions. It proposes Dr. MAS, which normalizes advantages per agent using individual reward statistics, achieving gains such as +5.6% avg@16 and +4.6% pass@16 on math benchmarks and +15.2% avg@16 and +13.1% pass@16 on search benchmarks while eliminating gradient spikes.
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.