SMRC-SD: Solving State-Reference Mismatch in Multi-Turn Agent Distillation
Junzhuo Liu · hf · 2026-08-10
When using successful trajectories for policy distillation in multi-turn agents, the student's execution state often mismatches the reference trajectory. To address this, researchers introduced SMRC-SD (State-Matched Routing and Contextualized Self-Distillation).
Core mechanisms:
- State-Matched Routing: At each turn, it verifies if the student's current state matches a supported state in the reference. Distillation is applied only at matched states, filtering out turns lacking local guidance.
- Contextualized Self-Distillation: For matched states, it constructs state-conditioned teacher context, grounding supervision in the actually reached state.
Across ALFWorld and WebShop, this method significantly boosts task success rates. Using the Qwen3-1.7B model, success rates improved from 0.746 to 0.865 on ALFWorld and from 0.574 to 0.693 on WebShop.
More from coding & agent
- Open Source Secure Browser MCP: Safe Browser Automation for AI Agents — Early_Resolution6932 · 2026-08-10
- AC2 Platform Enables Post-Training LLMs in Production, Avoiding Train-Test Mismatch — AccBalanced · 2026-08-10
- No-Code Tool to Generate PostgreSQL MCP Servers Instantly — Far_Thought2946 · 2026-08-10
- Won 5th Place in GPU Mode with Coding Agents, No CUDA Background — tokenbender · 2026-08-10
- Vibe Coding Trap: AI Can Patch Bugs Without Actually Fixing Them — Tired40s · 2026-08-10
- ReASearch: Single LLM Agent Outperforms Specialized Optimizers Across ML Workflows — _reachsumit · 2026-08-10