AI Science Tracker
EN

University of Bristol

1 discovery from 2026

Discoveries

Image scorers relied less on generated descriptions after reward training

Source check: Source checked. Not independent scientific confirmation.

Finding: Replacing an image’s generated reasoning with a description of another image reduced the rank correlation with human ratings by 0.548 for base 7B Qwen2.5-VL-Instruct and by 0.084 after score-reward training on KADID-10k. The tested 32B model retained greater sensitivity to the swap. The experiment measures dependence on the generated words for these image-scoring tasks.

Technology: Qwen2.5-VL-Instruct; reinforcement learning with verifiable rewards

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

AI contribution
Qwen2.5-VL-Instruct image scorers were trained with verifiable rewards and evaluated with reasoning descriptions swapped across images.
Test and result
The paper compares base and score- or rank-reward models on standard image-quality datasets, using donor-chain substitution and other interventions.
Limitation / still unknown
The experiment tests observable descriptions and task performance. It does not establish reasoning mechanisms for other tasks or model families.
Evidence sources and annotation context

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

Evidence sources: https://arxiv.org/abs/2610.33203v1 · https://arxiv.org/pdf/2610.33203v1

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 model setup and §5 intervention table.

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.