aifollow.news
Back Rohan Paul
Rohan Paul· @rohanpaul_ai · X· · Original publication time AI score54

Northwestern and Google paper proposes a way to improve agent harnesses for new tasksMachine translation

AI introduction

In the research, Harness2 runs an agent, has an editor model revise its prompts, tools, skills and memory, then reruns the task to keep useful changes. It stores lessons for later tasks. The source reports that one improvement step raised GLM-5’s WorkBuddy-Bench score by 11.5 points, while harness edits without test runs lowered coding scores from 73.4 to 63.2.

Original post · AI translation

@rohanpaul_ai
Translation · refer to the original

西北大学与谷歌的一篇新论文表明,智能体的运行框架无需标准答案,也能通过试运行修改并保留有效改动,在新任务上自我改进。

智能体的运行框架是围绕模型配置的提示词、工具、技能和记忆。手动调整这一框架很慢,而新的真实世界任务也没有可供对照调整的标准答案。

Harness2先运行智能体,让一个编辑模型提出运行框架的修正方案,然后重新运行任务,检验哪些修正改善了智能体的行为。每项任务中获得的经验都会进入一个经验库,用来指导下一项任务的修改。

经过一次改进,GLM-5在WorkBuddy-Bench上的得分提高了11.5分,超过了使用原版运行框架的Gemini 3.5 Flash。相反,如果修改运行框架时不进行这些试运行,编程任务得分会从73.4降至63.2。

应把智能体的运行框架视为需要针对每项任务修订的配置,并且只保留那些在实际运行中改变其行为的修改。

Source: Rohan Paul · x.com

Research
Found an error? Send a correction