Chain-of-thought (CoT) instructions can distort multiple-choice VLM evaluation when a scorer appends a reasoning cue but reads answer-label logits before the…
机构:上交
来源:arXiv 2609.29278 | AI4Papers 论文推荐平台