MAPD Distills Reasoning Protocols for Open Models

// 2h agoRESEARCH PAPER

MAPD Distills Reasoning Protocols for Open Models

Researchers introduced Multi-Agent Protocol Distillation (MAPD), a framework that transfers agentic search capabilities from proprietary LLMs into smaller open-source models via structured reasoning protocols. By feeding JSON-formatted exploration traces to a privileged student policy alongside reinforcement learning, MAPD significantly boosts QA benchmark performance for models like Qwen3-1.7B and Qwen3-4B.

// ANALYSIS

Distilling structured intermediate protocols rather than raw token trajectories proves that teaching open-source models how to structure their search reasoning is far more effective than imitating proprietary model style.

–Solves heterogeneous distillation without needing access to hidden teacher logits or matching tokenizers.
–Prevents style drift and verbosity degeneration by isolating reasoning protocols from surface-level generation artifacts.
–Combines offline multi-agent trace extraction with privileged student policy distillation and sparse RL objectives.
–Delivers strong empirical performance, lifting success rates on smaller models like Qwen3-1.7B (39.4%) and Qwen3-4B (44.4%).

// TAGS

llmagentic-searchagentknowledge-distillationreinforcement-learningopen-sourceqwenmapd

DISCOVERED

2h ago

2026-07-28

PUBLISHED

2h ago

2026-07-28

RELEVANCE

8/ 10

AUTHOR

_akhaliq

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCE36m ago

OpenAI open-sources Codex Security CLI

OpenAI has open-sourced Codex Security CLI, a tool designed to help developers discover, validate, and remediate vulnerabilities directly within their local development environment and CI/CD pipelines. By combining local context analysis with AI threat modeling, Codex Security CLI allows engineering teams to scan local repositories, automate security checks in workflows like GitHub Actions, and track security findings over time.

UPDATE46m ago

BAI adds four frontier models to API gateway

BAI, an API platform offering unified access to large language models, has expanded its offerings with four new frontier AI models. The update aims to empower developers by shortening the time required to deploy cutting-edge AI capabilities into applications through a centralized gateway.

MODEL1h ago

OpenAI launches gpt-live-transcribe for real-time transcription

OpenAI has introduced gpt-live-transcribe, a speech-to-text model designed for low-latency streaming transcription over live connections. Returning text deltas in real time as audio arrives, the model supports instant multilingual language switching, keyword hints for custom terminology, and noise filtering for challenging environment conditions.