New IAPO method improves LLM agent credit assignment in multi-turn tasks

Researchers have developed Influence-Aware Policy Optimization (IAPO), a new method for training large language model agents that interact with users and tools over multiple turns. IAPO represents agent rollouts as influence-dependency graphs, using user and tool observations to route advantages and improve credit assignment from sparse outcome feedback. Experiments with Qwen3-4B and Qwen3-8B models showed IAPO outperformed existing multi-turn reinforcement learning baselines on benchmarks like tau^2-Bench, UserBench, and AgentChangeBench, without sacrificing multi-turn function-calling capabilities. AI

影响 Enhances training of multi-turn LLM agents by improving credit assignment from sparse feedback.

排序理由 The cluster contains an academic paper detailing a new method for LLM agents. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.LG 阅读 →

• AgentChangeBench
• arXiv
• BFCL-v4 Multi-Turn
• Influence-Aware Policy Optimization
• Qwen3-4B
• Qwen3_8B
• tau^2-Bench

• 论文
• 模型发布

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →