多智能体强化学习何时能改善LLM工作流？工作流、规模与策略共享权衡

摘要

多智能体LLM工作流通过专业化角色分配推理路径以提升最终任务精度，但使用强化学习联合训练这些角色时，其不稳定性在机理上尚不明确。我们研究了端到端RL训练多智能体LLM工作流相比基模型的提升效果，比较了共享策略训练（所有角色更新同一策略）与隔离策略训练（各角色拥有独立参数）两种方案。实验矩阵涵盖Eval-Opt、Voting和Orch-Workers三种工作流，数学与代码两类任务，以及三个模型规模（0.6B、1.7B、4B）。研究发现：多智能体RL通常优于基模型，但增益同时取决于工作流、任务和规模，而非仅由策略共享决定。隔离策略训练往往能达到更高的峰值精度，但更频繁地遭遇终端精度悬崖；而共享策略训练并未消除失败，只是将失败重新分布为性质不同的模式。我们进而通过工作流拓扑与策略路由引发的角色级梯度动力学解释了其中最显著的模式：在隔离策略下，共享提示的并行同角色智能体会放大各角色梯度，导致Voting和Orch-Workers工作流出现终端退化；在共享策略下，非对称的逐步梯度质量导致共享策略被主导角色捕获，从而在任务与工作流维度产生不同的失败特征。综上，实证图谱及其内在机制表明，策略共享并非提供均匀稳定性，而是将训练压力导向不同通道，使其成为需权衡工作流与任务条件的设计选择。

English

Multi-agent LLM workflows route inference through specialized roles to lift end-task accuracy, but jointly training those roles with reinforcement learning is unstable in ways that are poorly understood. We study when end-to-end RL training of multi-agent LLM workflows improves over their base models, comparing Shared-Policy training, where all roles update one policy, with Isolated-Policy training, where each role has its own parameters. Our experimental matrix spans Eval-Opt, Voting, and Orch-Workers workflows, math and code tasks, and three model scales (0.6B, 1.7B, 4B). We find that multi-agent RL usually improves over base models, but gains depend jointly on workflow, task, and scale, not on policy sharing alone. Isolated-Policy tends to reach higher peak accuracy yet more often falls off a terminal accuracy cliff, while Shared-Policy training does not eliminate failure; it redistributes failure into qualitatively different patterns. We then explain the strongest of these patterns through role-level gradient dynamics induced by workflow topology and policy routing: under Isolated-Policy, parallel same-role agents on shared prompts amplify per-role gradients and drive terminal degradation in Voting and Orch-Workers workflows; under Shared-Policy, asymmetric per-step gradient mass causes the shared policy to be captured by the dominant role, producing different failure signatures by task and workflow. Together, the empirical map and its underlying mechanisms show that policy sharing routes training pressure through different channels rather than offering uniform stability, making it a design choice with workflow- and task-conditional tradeoffs.