Multi-Agent Protocol Distillation Bridges Distribution Gap in Agentic Search

A new paper on arXiv (2607.24280) introduces Multi-Agent Protocol Distillation (MAPD), a framework combining knowledge distillation and reinforcement learning to improve agentic search in large language models. Agentic search interleaves multi-step reasoning with retrieval for knowledge-intensive tasks, but outcome-based RL provides only sparse supervision. MAPD addresses the heterogeneous distillation problem by using a structured, style-normalized protocol as an intermediate representation, enabling distillation from proprietary models despite hidden logits and mismatched tokenizers. The framework employs an offline multi-agent setup to bridge the distribution gap between proprietary and open-source models, avoiding superficial stylistic artifacts from raw trajectory imitation. The paper proposes MAPD as a joint distillation and RL framework that densifies supervisory signals without requiring logit access.