SciFigPlag-Bench: Benchmark Challenges LLMs on Figure Plagiarism

anshulkundaje · x · 2026-08-19

Frontier LLMs still struggle to detect duplicated figures in academic journals, a challenging computational task, especially when plagiarism spans multiple articles. Researchers introduced SciFigPlag-Bench, a new benchmark with 2,582 positive and 2,541 negative pairs. It supports four diagnostic tasks: pairwise detection, source attribution, hierarchical reuse-type classification, and reuse correspondence localization. Experiments reveal persistent challenges in fine-grained provenance reasoning and spatial correspondence for current vision-language models.

Original post →

More from Multimodal

Multimodal channel →