From Proprietary to Open-Source: Bridging the Distribution Gap via Multi-Agent Protocol Distillation in Agentic Search
Junlin Liu, Jiangwang Chen, Zixin Song, Shuaiyu Zhou et al. · arXiv · 2026
Abstract
Teaching an open model to imitate a closed one is awkward: the closed model's logits are hidden, and copying its raw text traces mostly transfers writing style rather than reasoning. This paper inserts a structured intermediate — a JSON protocol carrying the task type, a plan, and grounding facts, produced offline by a multi-agent system — and trains the student against that. The setting is agentic search, where outcome-only reinforcement learning gives very sparse feedback.
Why it matters
A practical route for teams that want frontier-model behaviour from a model they can host themselves. The protocol-as-intermediate trick is the transferable idea; search is just where it gets demonstrated.
https://arxiv.org/abs/2607.24280