🛰️ Daily AI Frontier
‹ back to 2026-09-09

Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails

Research LLM Agents

Ranking

Overall 85
Content 95
Popularity 61

Observed public metrics from 1 member.

Representative image for Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails

Merged summary

TL;DR - This paper finds that fine-tuning weaker agent models to imitate complete expert trajectories can undermine their compatibility with an evolved agent harness. Correcting only failed turns in the weaker model’s own rollouts preserves its planning style and combines harness-evolution gains with model adaptation.

  • Across seven enterprise agent tasks, full-trajectory imitation reduced performance by 4–30 points for Qwen3-Coder and Gemma 4 under evolved harnesses.
  • Expert imitation transferred knowledge and increased scaffold usage but induced planning strategies the weaker models could not execute reliably.
  • The proposed on-policy pipeline uses a meta-level MLE agent to locate a failing turn, then asks an expert model to rewrite only that turn.
  • Localized correction maintained model–harness fit while helping economical, domain-specific agents narrow the performance gap.

Sources (1)

Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails

arXiv cs.AI Zhou Yu, Bin Bi, Shiva Kumar Pentyala, Shubham Mehrotra, Sougata Chaudhuri, Shilpa Bhagavath, Zeyuan Chen, Ran Xu, Phil Mui, James Zhu, Sitaram Asur 2026-09-08 arXiv:2609.09134
Public signals Hugging Face upvotes 11
Providers: Hugging Face · Upvotes 11 OpenAlex · N/A Publisher · N/A Semantic Scholar · N/A X · N/A Fetched 2026-09-25 14:22:29.676377 UTC

TL;DR - This paper finds that fine-tuning weaker agent models to imitate complete expert trajectories can undermine their compatibility with an evolved agent harness. Correcting only failed turns in the weaker model’s own rollouts preserves its planning style and combines harness-evolution gains with model adaptation.

  • Across seven enterprise agent tasks, full-trajectory imitation reduced performance by 4–30 points for Qwen3-Coder and Gemma 4 under evolved harnesses.
  • Expert imitation transferred knowledge and increased scaffold usage but induced planning strategies the weaker models could not execute reliably.
  • The proposed on-policy pipeline uses a meta-level MLE agent to locate a failing turn, then asks an expert model to rewrite only that turn.
  • Localized correction maintained model–harness fit while helping economical, domain-specific agents narrow the performance gap.
item →