ZICQ
中 Log in / Sign up
Newsroom Research & Papers #Multiwheel Smart #Knowledge distillation #Status Match #SMRC-SD

SMRC-SD: State-Matched Routing and Contextualized Self-Distillation for Multi-Turn Agents

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published: · 9 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:A new arXiv study introduces SMRC-SD, a framework that addresses state-reference mismatch in privileged distillation for multi-turn agents. By applying distillation only at matched states and constructing state-conditioned teacher context, SMRC-SD significantly outperforms unconditional full-path distillation on ALFWorld and WebShop, improving task success rates from 0.746 to 0.865 and 0.574 to 0.693 respectively with Qwen3-1.7B. This work offers a new paradigm for efficient training of interact


Background

In multi-turn interactive agent training, privileged distillation provides dense supervision by allowing a synchronized teacher to re-score the student's response at every turn. However, in interactive environments, the student's preceding actions continually change the execution state, potentially reaching states not covered by the reference trajectory, causing state-reference mismatch. Indiscriminate privileged distillation makes the reference unreliable.

Core Method

SMRC-SD (State-Matched Routing and Contextualized Self-Distillation) explicitly determines when and how a privileged trajectory should guide an on-policy student. Key mechanisms include:

  • State-Matched Routing: At each turn, verify whether the student's current state matches a supported state along the reference trajectory; apply distillation only at matched states, filtering out turns lacking locally compatible guidance.
  • Contextualized Self-Distillation: For each matched state, construct state-conditioned teacher context from the successful trajectory, grounding supervision in the actual state reached.

Experimental Results

On ALFWorld and WebShop benchmarks, SMRC-SD consistently outperforms unconditional successful full-path distillation. With Qwen3-1.7B, task success improves from 0.746 to 0.865 on ALFWorld and from 0.574 to 0.693 on WebShop. Ablations confirm that both selecting locally supported turns and constructing state-compatible teacher context contribute to the gains.

Technical Highlights

  • Addressing State Mismatch: State matching filters out incompatible guidance.
  • Contextualized Supervision: State-conditioned teacher context aligns supervision with actual states.
  • Significant Gains: Large improvements across benchmarks; code open-sourced.

Impact and Recommendations

This research offers a more reliable distillation method for multi-turn agents, especially in interactive environments. Developers can adopt state-aware supervision to improve generalization. The open-source code facilitates reproduction and further research.

References

  • Paper: https://arxiv.org/abs/2608.05219
  • Code: https://github.com/liujunzhuo/SMRC-SD
— END —

Tags: #Multiwheel Smart #Knowledge distillation #Status Match #SMRC-SD

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…