DocVLM-Fingerprint: Document VLM Faithfulness Under Visual Perturbations
Abstract DocVLM-Fingerprint studies a narrow failure mode in document visual question answering (VQA): a vision-language model (VLM) can return a plausible short answer while its accompanying claim is unsupported by the annotated evidence. This work implements a reproducible evaluation pipeline for generated chart, optical character recognition (OCR) document, and scientific figure tasks. The pipeline includes four visual perturbations, open VLM clients, answer scoring, claim-support labels, plots, and failure records. The main run covers 120 generated examples, 480 perturbed variants, and three open models. Small clean subsets of ChartQA and DocumentVQA, together with a Qwen-only CharXiv perturbation supplement using real charts, provide external checks. The evidence supports a controlled diagnostic study of failure patterns, not a general ranking of VLMs.
Authors
- Samarth Kukreja (ORCID: https://orcid.org/0009-0008-5494-9138)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-17
- DOI
- https://doi.org/10.21203/rs.3.rs-11757192/v1
- Primary Topic
- Multimodal Machine Learning Applications
- Type
- preprint