PapersBuilders10:42Chinese contentHSI框架:让冻结LLM通过分层自改进进化任务脚手架想用冻结模型提升agent性能?HSI让模型自己改写执行脚手架,在BabyAI等任务上涨了39个点,代码已开源,值得一看。#HSI#DeepSeek-V4-Flash-Preview#BALROG#agentsaarXiv: DeepSeek@Tailin ZhouSourceRead laterReadWorth followingUsefulFollow HSI