← News

Models & toolsSources: arXiv · 18 Nov 2026

Cited regions in small video-language models often do not drive answers, counterfactual test finds

Researchers blacked out the exact regions models cited in their reasoning chains and measured how many correct answers changed, exposing limits of standard grounding checks.

In plain words

Some small video-understanding models explain answers by citing the frames and boxes they looked at. The authors tested whether those citations actually matter by erasing them and seeing if answers change. Often the answers stayed correct, meaning the cited regions were not truly driving the result.

What was announced

A grounded reasoning chain from a video-language model names a frame span and a box at every step, so a reader can verify where the model looked. The arXiv paper argues that verifiability is not the same as dependence: an answer can pass a grounding check while ignoring the cited region. To measure this, the authors intervene directly, blacking out the regions a model cited in its own chain and comparing the flip rate of correct answers against blacking out same-sized boxes elsewhere on the same frames. The paired difference, which they call the cite gap, requires no annotation.

Applied to 2-3B video-language models trained on exact CLEVRER chains and pseudo-labelled NExT-QA chains, the cite gap was 0.089 on CLEVRER but only 0.003 on NExT-QA with the same weights. Pooled over models, masking cited boxes flipped 2.2 times as many correct answers as the control on CLEVRER and 1.1 times on NExT-QA, even among answers that depend on the video at all, while the boxes passed a grounding check on both datasets. Masking the cited entity on every frame flipped 1.9 times as many answers as its control. The authors note the NExT-QA chains were supervised to cite two frames of entities visible on most frames, so the weak result matches the supervision; their point is that the usual check cannot detect it. A citation can be verifiable and relevant without the answer depending on it, they conclude, and tests of grounded reasoning should remove what a chain names, not only where it points.

What it does not show

  • The evidence is the paper's abstract only; full methods, model details and results beyond the excerpt are not available for verification.
  • The findings cover 2-3B video-language models on CLEVRER and NExT-QA and do not establish how the cite gap behaves on other model sizes or datasets.
  • The paper reports its own measurements; no independent confirmation is implied.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.