Researchers from Hugging Face and other institutions have introduced Multi-Agent Protocol Distillation (MAPD), a novel framework designed to transfer the reasoning capabilities of proprietary large language models (LLMs) to open-source models for agentic search tasks. Agentic search involves interleaving multi-step reasoning with retrieval to solve knowledge-intensive queries, but optimizing such systems with outcome-based reinforcement learning (RL) often provides only sparse supervision.
Knowledge distillation from advanced proprietary models can offer denser guidance, but conventional methods face challenges: logit-matching is blocked by hidden logits and mismatched tokenizers, while direct imitation of natural language trajectories tends to copy superficial stylistic artifacts rather than core reasoning skills. MAPD addresses this heterogeneous distillation problem by using a structured, style-normalized protocol as an intermediate representation.
The framework employs an offline multi-agent system (MAS) that decomposes each query, retrieves supporting evidence, repairs failed searches, and converts the exploration trace into a JSON protocol containing the task type, reasoning plan, and extractive grounding facts. During training, this protocol is provided only to a privileged branch of the student policy, whose token distributions supply a dense distillation signal alongside the sparse RL objective.
Extensive evaluations across seven QA benchmarks show that MAPD consistently outperforms competitive distillation and RL methods. It achieves average success rates of 39.4% on Qwen3-1.7B and 44.4% on Qwen3-4B. Crucially, the framework generalizes robustly across diverse proprietary teachers while effectively mitigating style drift and verbosity degeneration in the student policy.