A benchmark score is commonly read as a proxy for model capability. We study the inverse problem: can a model be driven to a target score, and do its failure…
机构:中科院
来源:arXiv 2611.15568 | AI4Papers 论文推荐平台