UCSB and Microsoft's ScholarEvolve lets AI agents read papers to rewrite their own harness
Researchers at UC Santa Barbara and Microsoft have proposed ScholarEvolve, a system that automates reading research papers, implementing their methods as code and evaluating the resulting changes to an AI agent's harness, the software layer that surrounds a model. It splits the harness into five modules—Tools, Context, Skills, Memory and Workflow—and uses a Research Agent and a Coding Agent to retrieve work from sources such as arXiv, group approaches by mechanism through topic modeling and test candidate combinations. In author-reported results, ScholarEvolve raised Qwen3.5-27B's completion rate on AppWorld Challenge from 49.6% to 63.6% without changing the model's parameters, and also reported gains for GPT-5.4-mini on AppWorld Challenge and τ²-Bench Telecom; the figures have not been independently verified.
The work targets the harness layer around a model rather than model weights, reporting that automatically importing methods from research literature can improve agent task completion, though the benchmark numbers are author-reported.
机器之心 2026-10-21 12:01 北京
AI Agent 能否像研究人员一样，从论文中学习？
最近两年，大语言模型的能力不断提升，围绕模型构建的 AI Agent 也逐渐能够处理复杂的任务。从调用工具、操作软件，到规划和执行多步骤工作流，Agent 从处理简单的问答，走向能够独立解决长程复杂工具调用的任务。与此同时，每天也都有很多新的 agent 领域的 preprint 公开在 arXiv 中。
但一个现实问题随之出现：当 Agent 遇到能力瓶颈时，我们该如何让它变得更好？
#ScholarEvolve#UCSB#Microsoft#Qwen3.5-27B#GPT-5.4-mini#AppWorld#τ²-Bench#Meta Harness