From Proprietary to Open-Source: Bridging the Distribution Gap via Multi-Agent Protocol Distillation in Agentic Search
MAPD trains smaller open models from proprietary search traces without copying their surface style.
The paper proposes a structured JSON protocol that captures task type, reasoning plan, and grounded facts from an offline multi-agent search process. That protocol feeds a privileged student-policy branch during training, adding dense distillation alongside sparse reinforcement learning. Across seven QA benchmarks, it reports average success rates of 39.4% for Qwen3-1.7B and 44.4% for Qwen3-4B. The authors say the approach works across different proprietary teachers while reducing style drift and verbosity degeneration. HF Daily Papers' note
The paper proposes a structured JSON protocol that captures task type, reasoning plan, and grounded facts from an offline multi-agent search process. That protocol feeds a privileged student-policy branch during training, adding dense distillation alongside sparse reinforcement learning. Across seven QA benchmarks, it reports average success rates of 39.4% for Qwen3-1.7B and 44.4% for Qwen3-4B. The authors say the approach works across different proprietary teachers while reducing style drift and verbosity degeneration. HF Daily Papers' note
score 5