Agentic ESOpt：以最小 GPU 需求微调长时程 LLM 智能体
原标题：Paper page - Agentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Requirements
AI 摘要
Agentic ESOpt 提出使用进化策略（ES）对长时程 LLM 智能体进行全参数微调，仅需推理级 GPU 内存，避免了强化学习（RL）中反向传播的高成本与长时程信用分配难题。该方法在 WebArena-Lite 上将 Qwen-3.5-27B 的 No Skill 基线提升 6.69%，并在 28/36 个测试时自动启发式设计设置中优于匹配基线。
正文节选
Agentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Requirements Abstract Agentic ESOpt uses evolution strategies for scalable full-parameter fine-tuning of long-horizon LLM agents via trajectory-level reward-weighted updates and parameter-context co-evolution. Reinforcement Learning (RL) has been promising in single-turn LLM fine-tuning. However, long-horizon agentic reasoning introduces increasingly branching interactions and sparse rewards, exposing several limitations of RL: