Harness-Zero: Harness Distillation via Agent-as-Harness
AI Digest
代理蒸馏支架迁移模型优化行为恢复
Harness-Zero通过代理作为支架实现蒸馏，解决模型部署中支架依赖问题，实验显示其显著提升任务成功率并恢复基础模型缺失行为。
Harness-Zero enables distillation via agent-as-harness, overcoming deployment dependency issues. Experiments show 44.3% task success rate and 82.3% behavior recovery across domains.
Key points
- 传统支架依赖导致模型性能受限，需选择性部署专用支架 Traditional harness dependency limits model performance with selective deployment
- Harness-Zero通过代理在目标支架空间修正响应实现行为迁移 Harness-Zero corrects responses in target harness space via agent guidance
- 实验验证该方法在知识工作、工具使用等场景均优于代码支架 Experiments validate superior performance across knowledge work and tool use
- 移除专用支架后模型表现反超带支架的基准模型 Model outperforms baseline with harness removal at deployment
- 成功恢复基础模型中缺失的82.3%行为模式 82.3% behavior recovery from base model limitations
Takeaway: 通过代理蒸馏实现支架无关的模型优化，突破传统部署限制 / Agent-based distillation enables harness-agnostic model optimization
Why it matters 提供新型模型优化范式，对实际部署中的支架依赖问题有直接解决价值
View original ↗ Back to hot list
This page is an aggregated digest from arxiv; content and hot-score data come from public sources. Copyright belongs to the original authors. We link to originals with nofollow and never republish full text.