AI Science Tracker
EN

Controlled stimuli reveal motion information in video Transformer features

University of Chicago; Toyota Technological Institute at Chicago · 23 Oct 2026 · computer science

Discovery

Controlled stimuli reveal motion information in video Transformer features

Source check: Source checked. Not independent scientific confirmation.

Finding: Under controlled motion stimuli, a statistical test flagged 28.7% to 92.1% of final-layer coordinate axes across the tested video Transformers, with zero flags in the frame-by-frame DINOv2 control. The study measured co-variation with specified motion profiles. Activation interventions gave mixed evidence that the flagged features drove model decisions.

Technology: VideoMAE; V-JEPA2; TimeSformer; DINOv2; sparse autoencoders

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
Pretrained video Transformers provided learned features whose activation covariances were tested under controlled motion stimuli.
Test and result
The authors evaluated final-layer coordinates and learned sparse features on designed visual stimuli, with multiple-comparison thresholds and an image-encoder comparison.
Limitation / still unknown
Positive tests indicate some temporal dependence under these stimuli and definitions. Rates do not measure pure temporal-feature prevalence or identify a mechanism.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.33353v1 · https://arxiv.org/html/2610.33353v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: v1 Methods §3–4, Table 1 and Appendix F.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.