AIAny
Icon for item

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

Bridges the proprietary-to-open-source gap in agentic search by converting multi-step retrieval and reasoning traces into a structured, style-normalized JSON protocol and using it for joint distillation + RL. Produces denser supervision that improves student success rates while reducing style drift.

Introduction

Most agentic search setups struggle to transfer reasoning competence from strong proprietary teachers to smaller open models: hidden logits, tokenizer mismatches, and stylistic artifacts either block direct logit-matching or produce superficial imitation. The core insight of this paper is that a structured intermediate—an extractive, style-normalized JSON protocol—can carry the teacher’s reasoning plan and grounding facts without transferring verbose stylistic signals, enabling dense distillation signals alongside sparse RL rewards.

Key Findings
  • Multi-Agent Protocol Distillation (MAPD) uses an offline multi-agent system to decompose queries, retrieve evidence, repair failed searches, and emit a JSON protocol containing task type, plan, and grounding facts — this protocol serves as a privileged dense target during student training, so the student learns procedure and grounding rather than surface style.
  • Joint training (protocol-guided distillation + RL) consistently outperforms standard distillation and pure RL baselines: reported average success rates reach 39.4% on Qwen3-1.7B and 44.4% on Qwen3-4B, showing robustness across different proprietary teachers and student sizes.
  • The protocol abstraction mitigates verbosity and style drift that arise from direct trajectory imitation, meaning students generalize reasoning behavior rather than mimic teacher phrasing; this makes the method more portable across teacher models with different styles.
Who it's for and trade-offs

Great fit if you need to transfer complex, retrieval-augmented multi-step reasoning from closed, high-capacity teacher models into smaller or open-source student policies while avoiding stylistic overfitting. It is especially useful when teacher logits are inaccessible or tokenizers differ. Look elsewhere if you require fully online distillation from a live teacher (MAPD relies on an offline MAS to produce protocols) or when computational budget forbids running the retrieval/repair pipeline needed to generate dense protocols.

Where it fits

MAPD sits between black-box teacher imitation and reward-only RL: it provides dense intermediate supervision that is model-agnostic (works with multiple proprietary teachers) and complements RL objectives to improve long-horizon, retrieval-heavy QA and agentic search tasks.

Method sketch

The pipeline converts multi-agent exploration traces into JSON protocols; during training a privileged branch consumes protocols to produce token distributions that supply a dense distillation loss, while the full student policy is optimized jointly with an outcome-based RL objective. Evaluations span seven QA benchmarks and cross-teacher settings to validate generalization and robustness.

Information

  • Websitearxiv.org
  • OrganizationsBeijing Institute of Technology, East China Normal University, University of Science and Technology of China, Tsinghua University, University of Chinese Academy of Sciences
  • AuthorsJunlin Liu, Jiangwang Chen, Zixin Song, Shuaiyu Zhou, Chunji Lv, Hank Wu, Kailin Jiang, Jinyang Wu, Bohan Yu, Chenxi Zhou
  • Published date2026/07/27

More Items

Drives long‑horizon desktop agents by reading and manipulating program state (files, DOM, backends) instead of relying on screenshots. The main agent uses code for actions and structural verification while a lightweight GUI subagent handles rare screenshot-click steps, improving success rates and lowering per-task cost versus screenshot-only approaches.

Provides a unified survey of progress-reward modeling for robotic learning, detailing interfaces, modeling techniques, and evaluation practices. Organizes the literature into three perspectives—interface, model internals, and data/benchmarks—and highlights limitations and open problems. Useful for researchers designing rewards for long-horizon or sparse-reward robotic tasks.

Lets canvas-native agents plan, generate, edit, and organize long-horizon multimodal creative projects by representing artifacts, versions, and actions as typed canvas nodes and links. Uses a three-layer design (canvas state, protocol bridge, agent runtime) so agents act within an inspectable, editable project state.