← 研踪 PaperTrailarXiv:2611.21940 · v1
Does the Model Use What It Cites? Counterfactual Tests of Grounded Reasoning Chains in Small Video-Language Models
首次提交 2026-11-17 06:03
版本更新 2026-11-17 06:03
雷达观测 2026-11-18 19:43
摘要
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.
作者备注:24 pages, 4 figures, 27 tables. Code: https://github.com/JugalGajjar/CausalSight
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-17 06:03 · 最新观测版本
Does the Model Use What It Cites? Counterfactual Tests of Grounded Reasoning Chains in Small Video-Language Models
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.
查看此版本解读 →查看此版本原文 ↗