ReMMDBench (ReMMD-Agent): leaderboard
Multimodal misinformation verification on 500 real multilingual posts with several images each, labelled with a five-level veracity verdict and eight distortion types; ReMMDBench (Dang et al., 2026). The model runs as the backbone of the authors' ReMMD-Agent verifier, which parses atomic claims and image bindings, retrieves web, image and social evidence and reuses it through an evidence memory. Macro-F1 in percent over the five veracity verdicts; higher is better.
Metric: Veracity macro-F1 (%; five-level verdict from true to false on 500 multilingual multi-image posts; backbone inside the ReMMD-Agent verifier with Serper web, image and social search). Source: arxiv.org. Saturation forecast: Around July 2028. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 39.12 | #172 |
| 2 | Qwen 3.5 9B | 37.18 | #803 |
| 3 | Gemma 4 31B | 33.76 | #393 |
| 4 | Qwen 3.6 27B | 30.07 | #339 |
| 5 | Qwen 3.5 4B | 28.58 | #1180 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=remmdbench-remmd-agent · How It Works · Data refreshed daily, snapshot 2026-11-12.