AI Science Tracker
EN

MADI

1 discovery from 2026

Discoveries

MADI

Near-duplicate image choices transferred preferences between vision-language models

Source check: Source checked. Not independent scientific confirmation.

Finding: Vision-language models used as curators chose among near-duplicate image variants for a student model’s training set. With underlying photos and human-written captions held fixed, students trained on selections from owl-preferring curators gave greater preference to owls in an image-free likelihood test across four tested model variants. The result concerns experimentally induced curator preferences under specified data and training conditions.

Technology: LLaVA-1.5; Qwen2.5-VL; Qwen3-VL; Gemma 4

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

AI contribution
Pretrained curator models selected one variant per photo using caption likelihood. Student models initialized from the same checkpoint as their curator learned from those image–caption pairs and were tested independently of the source images.
Test and result
Matched student groups saw image variants selected by target or neutral curators. The core comparison holds underlying source photographs and human-written captions fixed; supplementary follow-up timing is not used to establish the October 20 result.
Limitation / still unknown
Initial curator preferences were experimentally induced, and transfer varied by model and training conditions. The primary readout measures relative answer likelihood in specified prompts. The result does not establish ordinary production curation transfers broader preferences.
Evidence sources and annotation context

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

Evidence sources: https://arxiv.org/abs/2610.27943 · https://arxiv.org/html/2610.27943v1 · https://arxiv.org/html/2610.27943v2

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: Original arXiv version metadata, v1 abstract and controlled optical preference-transfer result and matched source-bank comparison, plus analogous v2 Methods §§2–3 and control appendix..

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 the interactive profile for the timeline, or browse all organizations and model families.