AI Science Tracker
EN

Activation directions shifted both image and text generation

Mila and McGill University; Carnegie Mellon University; MIT CSAIL; Northeastern University · 26 Oct 2026 · computer science

Discovery

Activation directions shifted both image and text generation

Source check: Source checked. Not independent scientific confirmation.

Finding: In UniAR, Emu3.5 and Liquid, some activation directions derived from text also changed image outputs, and some image-derived directions changed text. Emotion transferred in both directions more consistently than size or left–right relations. Transfer also depended on model and layer.

Technology: UniAR, Emu3.5 and Liquid; activation interventions

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

AI contribution
Activation directions derived from learned image or text representations were applied to generation in the other modality to measure how it changed.
Test and result
The study perturbed the shared model representations and compared outputs with unperturbed controls for eight concepts. Three reviewers each rated 320 separate randomized output pairs; their majority labels and the automated judge had agreement of κ=0.62.
Limitation / still unknown
The concept sets include generated examples. The response varies by model, concept and intervention strength, and transferring a direction does not establish that the underlying computations are identical.
Evidence sources and annotation context

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

Evidence sources: https://arxiv.org/abs/2610.36269v1 · https://arxiv.org/html/2610.36269v1

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: §§3.1–4 and Appendix A.8, including held-out blinded human ratings.

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 3/5. 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.