MM-IFEval-Pro (Without Hijacking): leaderboard
MM-IFEval-Pro (Xiao et al., 2026) scored without its visual instruction-hijacking cases, so only the ordinary image-grounded instruction-following items count, each with several constraints checked by deterministic rule-based verifiers. Mean per-sample constraint-satisfaction rate in percent; higher is better.
Metric: Accuracy (%; mean per-sample constraint-satisfaction rate x 100 over the bilingual Chinese and English test set, each constraint checked by a deterministic rule-based verifier (about 3 constraints per sample); evaluated without the visual instruction-hijacking cases). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 8B Instruct | 81.21 | #982 |
| 2 | Qwen 3.5 35B A3B | 74.6 | #480 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mm-ifeval-pro-without-hijacking · How It Works · Data refreshed daily, snapshot 2026-11-12.