AEGIS (Academic Image Forensics) - Manipulation Classification: 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 naming the kind of edit made to a highlighted region, in percent; higher is better.

Metric: Macro-F1 (%) over three classes, naming a highlighted edited region, given the original caption, as an insertion, removal or alteration, 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 2029. 26 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.148.34#240
2Gemini 3 Pro (Preview)45.42#119
3Doubao-Seed-1.6Thinking41.85
4Llama 4 Maverick40.18#1110
5GPT-4.140.05#505
6Claude Sonnet 4.539.7#261
7LLaVA-OneVision-72B39.42#1771
8Gemini 2.5 Flash37.28#463
9Seed-1.637.08#469
10Qwen 3 VL Plus36.89#512
11Qwen-VL-Max (AEGIS checkpoint unspecified)36.76
12Qwen 2.5 VL 72B Instruct35.96#841
13O4 MiniHigh35.54#368
14Gemma 3 27B (IT)34.76#1324
15Ministral 3 14B33.89#1870

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