SciFigPlag-Bench: leaderboard
SciFigPlag-Bench (2026) tests vision-language models on scientific figure plagiarism: detection, source attribution, reuse-type classification and localization of reused regions over real and synthetic figure pairs. The overall score is the mean of the four task accuracies in percent, higher is better.
Metric: Overall score (%; mean of pairwise-detection accuracy, source-attribution accuracy, reuse-type joint accuracy and paired reuse-localization accuracy; SciFigPlag-Bench pairs of source and suspicious scientific figures (2,582 plagiarized pairs from documented real cases and taxonomy-guided synthesis, 2,541 visually similar negatives); the same task-specific prompts for every model, temperature 0). Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 83 | #164 |
| 2 | Gemma 4 31B (IT) | 82.7 | #342 |
| 3 | Qwen 3.6 35B A3B | 80.5 | #462 |
| 4 | GPT-5.4 | 80.1 | #123 |
| 5 | Qwen 3.5 35B A3B | 79.8 | #482 |
| 6 | Claude Sonnet 4.6 | 77.6 | #161 |
| 7 | Gemma 4 26B A4B (IT) | 53.5 | #549 |
| 8 | Mistral Small 4 | 37.3 | #1201 |
| 9 | Qwen 3.5 2B | 30.9 | #4810 |
| 10 | Gemma 3 4B (IT) | 28.6 | #3831 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=scifigplag-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.