AEGIS (Academic Image Forensics) - Forgery Scope Discrimination: leaderboard
AEGIS (2026) tests whether multimodal models can audit AI-generated or AI-edited images in academic papers, from figure panels across seven scientific image categories and forgeries made with 25 image generators. Macro-F1 for deciding whether an image is authentic, partly edited or fully generated, in percent; higher is better.
Metric: Macro-F1 (%) over three classes, telling whether an image is authentic, partly AI-edited or entirely AI-generated (answers of Not Sure, an offered abstention, are left out), on AEGIS academic images (seven categories, 39 subtypes; real panels and forgeries from 25 generative models under four forgery strategies), zero-shot minimal prompts, lossless PNG input; higher is better. Source: arxiv.org. Saturation forecast: Around May 2027. 26 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | GPT-4.1 | 59.98 | #505 | |
| 2 | Gemini 2.5 Flash | 53.78 | #463 | |
| 3 | Gemini 3 Pro (Preview) | 53.44 | #119 | |
| 4 | O4 Mini | High | 49.35 | #368 |
| 5 | GPT-5.1 | 46.16 | #240 | |
| 6 | Gemma 3 27B (IT) | 42.36 | #1324 | |
| 7 | llava-onevision-qwen2-7B-ov-chat | 39.94 | ||
| 8 | LLaVA-OneVision-7B | 38.76 | #2981 | |
| 9 | Qwen-VL-Max (AEGIS checkpoint unspecified) | 37.3 | ||
| 10 | Doubao-Seed-1.6 | Thinking | 37.13 | |
| 11 | Llama 4 Maverick | 37 | #1110 | |
| 12 | Qwen 3 VL Plus | 35.72 | #512 | |
| 13 | Seed-1.6 | 35.38 | #469 | |
| 14 | Doubao-1.5-thinking-vision-pro | 35.31 | #344 | |
| 15 | Qwen 2.5 VL 72B Instruct | 33.16 | #841 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=aegis-academic-image-forensics-forgery-scope-discrimination · How It Works · Data refreshed daily, snapshot 2026-11-21.