Multi-Agent Protocol Distillation Transfers Proprietary Agentic Search Behavior Into Qwen3-1.7B at 39.4% Success

arXiv 2607.24280·medium signalXBlueskyLinkedInCopy link

'From Proprietary to Open-Source' (arXiv 2607.24280, July 27) attacks the sparse-supervision problem in outcome-based RL for agentic search by inserting structured JSON protocols — task types, reasoning plans, grounding facts — as an intermediate representation distilled from proprietary models into open-source ones. MAPD decomposes queries, retrieves evidence, repairs failed searches, and converts exploration traces into those protocols, which then guide a privileged training branch and supply dense distillation signal alongside the sparse RL objective. Reported gains hold across seven QA benchmarks, reaching 39.4% success on Qwen3-1.7B and 44.4% on Qwen3-4B while suppressing style drift and verbosity degeneration — a credible recipe for moving agentic search capability onto models you can self-host.