Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

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

Source

arXiv

arxiv.org/abs/2607.24280v1

Read the source

Source

Plain-Language Brief

Researchers propose Multi-Agent Protocol Distillation (MAPD), a method to bridge the gap between open-source and proprietary language models in solving knowledge-intensive tasks. MAPD uses a structured protocol as an intermediate representation for dense guidance, addressing challenges of hidden logits and mismatched tokenizers in conventional distillation methods.

What the Document Says

  • Agentic search is used to solve knowledge-intensive tasks by interleaving multi-step reasoning with retrieval. Optimizing this with outcome-based reinforcement learning (RL) provides sparse supervision.
  • Knowledge distillation can supply denser guidance, and advanced proprietary models are promising teachers for distilling.
  • Conventional logit-matching is precluded by hidden logits and mismatched tokenizers in proprietary models, and raw natural language trajectory imitation transfers superficial stylistic artifacts instead of core reasoning competence.
  • The authors propose Multi-Agent Protocol Distillation (MAPD), a joint distillation and RL framework that uses a structured, style-normalized protocol as an intermediate representation.
  • An offline multi-agent system decomposes each query, retrieves supporting evidence, repairs failed searches, and converts the resulting exploration trace into a JSON protocol containing task type, reasoning plan, and extractive grounding facts.
  • During training, the protocol is provided only to a privileged branch of the student policy, whose token distributions furnish a dense distillation signal alongside the sparse RL objective.
  • Evaluations on seven QA benchmarks show that MAPD outperforms competitive distillation and RL, achieving average success rates of 39.4% on Qwen3-1.7B and 44.4% on Qwen3-4B.
  • MAPD generalizes robustly across diverse proprietary teachers while effectively mitigating the student policy from style drift and verbosity degeneration.

Key Concepts Mentioned

  • Agentic search
  • knowledge distillation
  • reinforcement learning
  • proprietary models
  • multi-agent protocol
  • JSON protocol

Claims Made by the Document

  • MAPD outperforms competitive distillation and RL.
  • MAPD generalizes robustly across diverse proprietary teachers.

Evidence Used

  • Extensive evaluations on seven QA benchmarks demonstrate that MAPD achieves average success rates of 39.4% on Qwen3-1.7B and 44.4% on Qwen3-4B.
  • The framework effectively mitigates the student policy from style drift and verbosity degeneration.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Agentic search enables large language models to solve knowledge-intensive tasks by interleaving multi-step reasoning with retrieval, yet optimizing this with outcome-based reinforcement learning (RL) provides only sparse supervision. Knowledge distillation can supply denser guidance, and advanced proprietary models with their strong reasoning capabilities are promising teachers. While distilling from proprietary models can densify this supervisory signal, conventional logit-matching is precluded by hidden logits and mismatched tokenizers, whereas raw natural language trajectory imitation transfers superficial stylistic artifacts rather than core reasoning competence. To address the heterogeneous distillation problem and bridge the distribution gap, we propose Multi-Agent Protocol Distillation (MAPD), a joint distillation and RL framework uses a structured, style-normalized protocol as an intermediate representation. An offline multi-agent system (MAS) decomposes each query, retrieves supporting evidence, repairs failed searches, and converts the resulting exploration trace into a JSON protocol containing the task type, reasoning plan, and extractive grounding facts. During training, the protocol is provided only to a privileged branch of the student policy, whose token distributions furnish a dense distillation signal alongside the sparse RL objective. Extensive evaluations across seven QA benchmarks demonstrate that MAPD consistently outperforms competitive distillation and RL, achieving 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 the student policy from style drift and verbosity degeneration.