ERT让视觉模型更易被大语言模型编辑
AI 总结
Timo Ropinski 和 Sebastian Kranz 发表了论文《Training Vision Models for LLM-Driven Weight-Space Programming》,其基准让五个大语言模型在重标注和纠错任务中编辑卷积网络与视觉 Transformer。论文提出 Edit-Ready Training(ERT),通过预演直接、闭式的权重编辑来训练模型;作者报告了超过2,600次编辑,并称 ERT 在全部评测基线中为 GPT-6 Astra 和 Claude Opus 5.5 取得最高平均分。
深度见解
这项研究通过具体方法和基准,直接考察专门训练能否让大语言模型更容易编辑神经网络权重。
Computer Science > Computer Vision and Pattern Recognition
arXiv:2611.26408 (cs)
[Submitted on 19 Nov 2026]
Title:Training Vision Models for LLM-Driven Weight-Space Programming
Authors:Timo Ropinski, Sebastian Kranz
View a PDF of the paper titled Training Vision Models for LLM-Driven Weight-Space Programming, by Timo Ropinski and 1 other authors
#Edit-Ready Training#GPT-6 Astra#Claude Opus 5.5#Waterbirds#CIFAR-100#PACS#Entity-30