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

Researchers introduced Multi-Agent Protocol Distillation, a hybrid distillation-plus-RL method for agentic search models.

It turns multi-agent search traces into a structured JSON protocol and uses dense guidance from proprietary teacher models.

On seven QA benchmarks, it beat prior distillation and reinforcement learning baselines.

The method improved performance on Qwen3-1.7B and Qwen3-4B while reducing sparse supervision and tokenizer mismatch issues.