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

  • 类型：arxiv
  • 标识：2609.09134
  • 链接：https://arxiv.org/abs/2609.09134
  • 主分类：evaluation
  • 形态：method
  • 被引：4
  • 被引来源：Semantic Scholar
  • S2被引：4
  • OpenAlex被引：0
  • 影响力被引：0
  • TLDR：An on-policy expert-correction pipeline is developed, automated by a meta-level MLE agent, that localizes the failing turn in the weaker model's own rollout and asks the expert to rewrite only that turn, which preserves the model's planning style and combines the gains of harness evolution and model adaptation.
  • OpenAlex ID：W7212074373
  • OpenAlex DOI：10.48550/arxiv.2609.09134
  • DOI：10.48550/arxiv.2609.09134
  • DOI来源：OpenAlex
  • 开放获取：green
  • 开放获取链接：https://arxiv.org/pdf/2609.09134
  • OpenAlex更新：2026-09-12
  • 副分类：agent
  • 待LLM分类：否
  • 标题中文：协同进化的 Harness 与模型：On-policy 修正助力弱模型在模仿失效处迎头赶上
  • TLDR中文：开发一条 on-policy 专家修正流水线，由元层级 MLE Agent 自动化，在弱模型自身的 rollout 中定位失败回合，并请专家仅重写该回合，从而保留模型的规划风格，融合 Harness 进化与模型适配带来的收益。
  • 来源文件：
  • /inbox/tom/_candidates/2026-09-10-agent-rag-longcontext-candidates.json
  • [S2 enrich]
  • [OpenAlex backfill]