From Proprietary to Open-Source: Bridging the Distribution Gap via Multi-Agent Protocol Distillation in Agentic Search

Researchers introduced MAPD, a hybrid distillation-plus-RL framework for agentic search training.

An offline multi-agent pipeline converts search traces into a normalized JSON protocol with task type, reasoning plan, and grounding facts.

That structured protocol provides dense supervision to a privileged student branch, avoiding reliance on hidden logits or direct text imitation.

MAPD outperforms prior distillation and RL methods across seven QA benchmarks, including open-source Qwen3 models.

The approach helps transfer search and reasoning behavior from proprietary teachers to open-source agents for knowledge-intensive tasks.