AEGIS (Academic Image Forensics): 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. Composite forensic index that combines the three classification tasks with tampering localization and penalizes over-localization, in percent; higher is better.

Metric: Normalized Forensic Index (%): harmonic mean of the macro-F1 scores of the three classification tasks and the correct localization accuracy of tampering pinpointing, times the square root of one minus the over-localization rate, 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 October 2027. 26 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.148.8#240
2Gemini 2.5 Flash47.02#463
3Doubao-Seed-1.6Thinking46.66
4Gemini 3 Pro (Preview)45.79#119
5GPT-4.143.31#505
6Qwen-VL-Max (AEGIS checkpoint unspecified)43.01
7O4 MiniHigh42.77#368
8Doubao-1.5-thinking-vision-pro42.39#344
9Seed-1.641.73#469
10Qwen 2.5 VL 72B Instruct38.71#841
11Gemma 3 27B (IT)37.55#1324
12Doubao-Seed-1.6-Flash36.03#1250
13Doubao-1.5 Vision Pro34.07#1026
14Qwen-VL-Plus (AEGIS checkpoint unspecified)33.89
15LLaVA-OneVision-72B33.81#1771

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 · How It Works · Data refreshed daily, snapshot 2026-11-21.