模型是否真的用了它引用的区域?小型视频语言模型中有据推理链的反事实检验
Does the Model Use What It Cites? Counterfactual Tests of Grounded Reasoning Chains in Small Video-Language Models
AI 导读
一项 arXiv 论文提出“cite gap”反事实测试:将 2-3B 视频语言模型推理链中引用的区域涂黑,与同帧同大小随机区域对比答案翻转率。在 CLEVRER 链上训练的模型 cite gap 为 0.089,涂黑引用框翻转正确答案的次数是对照的 2.2 倍;但在 NExT-QA 链上 cite gap 仅 0.003,为对照的 1.1 倍。
正文 · 原文
Abstract:A grounded reasoning chain cites a frame span and a box for every step, so a reader can check where the model looked. Whether the answer depends on the cited region is a different question. We test it by intervention: the regions a model cited in its own chain are blacked out, and the share of correct answers that flip is compared with blacking out boxes of the same size elsewhere on the same frames. The paired difference, the cite gap, needs no annotation. We apply it to 2-3B video-language models trained on exact CLEVRER chains and on pseudo-labelled NExT-QA chains. With the same weights, the cite gap is 0.089 on CLEVRER and 0.003 on NExT-QA. Pooled over models, masking the cited boxes flips 2.2 times as many correct answers as the control on CLEVRER and 1.1 times as many on NExT-QA, also among the answers that depend on the video at all, while the boxes pass a grounding check on both. Masking the cited entity on every frame, instead of the cited box, flips 1.9 times as many answers as its control. The NExT-QA chains were supervised to cite two frames of entities that stay visible on most frames, so this outcome is what the supervision predicts; the point is that the usual check on grounded chains cannot see it. A citation can be verifiable and relevant without the answer depending on it, and a test of grounded reasoning should remove what a chain names, not only where it points.
| Comments: | 24 pages, 4 figures, 27 tables. Code: this https URL |
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Machine Learning (cs.LG) |
| Cite as: | arXiv:2611.21940 [cs.CV] |
| (or arXiv:2611.21940v1 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2611.21940 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Jugal Gajjar [view email]
[v1]
Mon, 16 Nov 2026 22:03:19 UTC (952 KB)
来源:arXiv cs.AI · arxiv.org