FedAgent · Is Decentralized LLM Agent RL Robust to Heterogeneity?

FedAgent (Federated Agent Reinforcement Learning) trains LLM agents across distributed clients without sharing data. Its robustness is asymmetric: robust to task-level heterogeneity, worst-case fragile to environment-level heterogeneity, explained by the Input-Dynamics Asymmetry of task-augmented M…

关键词：federated learning, reinforcement learning, AI agents, LLM agents, GRPO, PPO, privacy-preserving machine learning, decentralized training, heterogeneity, WebShop, ALFWorld, input-dynamics asymmetry