AI captions added unsupported detail in the middle and more reference-unmatched content near the end
Source check: Source checked. Not independent scientific confirmation.
Finding: In long captions for 198 images, six AI models produced more reference-unsupported specifics in the middle. The mean share assigned to a residual reference-unmatched category rose from 10.3% in the first quarter to 28.1% in the last. This category need not mean that a sentence is objectively vague.
Technology: Qwen3-based claim extraction and verification; Qwen3-VL, InternVL3.5 and LLaVA captioners
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Qwen models extract and verify ordered atomic claims from human references and generated captions; human annotators validate a sample of judgments.
- Test and result
- Researchers divided captions into quarters, assessed claims against human-derived image rubrics, and checked a sample of verifier judgments with human reviewers.
- Limitation / still unknown
- A caption can include a true detail omitted from its reference. The reported pattern concerns this sample and reference-based labeling, with only sampled image-level checks.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.31111 · https://arxiv.org/html/2610.31111v1
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: arXiv v1 §3, Figure 1/Table 1, appendix comparison to 347 human rubric judgments and sampled visual recheck.
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 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.