Reproduce “Internalizing Multi-Agent Reasoning for Accurate and Efficient LLM-based Recommendation” on OpenResearch

About this paper

Yang Wu, Haoze Wang, Qian Li, Jun Zhang, Huan Yu, Jie Jiang

A framework called STAR internalizes multi-agent reasoning, including tool use and self-reflection, into a single Large Language Model for recommendation systems. This approach effectively combines semantic reasoning with collaborative signals, achieving performance often surpassing multi-agent teachers while significantly reducing inference latency.

Key insights

• Complex multi-agent reasoning, including planning, tool usage, and self-reflection, can be effectively internalized into a single LLM, achieving comparable or superior performance to its multi-agent teacher.
• The distillation process, particularly with Group Relative Policy Optimization, effectively filters noise and prevents error propagation, leading to a more robust unified policy in the student model.
• Verbalizing implicit collaborative signals from graph structures into natural language evidence, pre-computed offline, allows LLMs to explicitly reason over behavioral patterns efficiently.

Summary by alphaXiv.