VSysBench - Task Accuracy: leaderboard
VSysBench (2026) pairs MM-Vet v2 image questions with system messages that impose a constraint, from output format and persona to vision-grounded rules, and in half of the cases adds a user message that tries to override it; a GPT-5-mini judge scores the answer and the compliance. Score is the judged answer accuracy while the constraint applies, over both halves, in percent; higher is better.
Metric: Task accuracy (%; mean judged answer score against the MM-Vet v2 ground truth while the system constraint applies, both halves; VSysBench: 2,258 verified MM-Vet v2 image questions, each under a system-message constraint from 22 sub-categories in 5 categories, once with an aligned and once with a conflicting user message (4,516 instances); default inference settings; GPT-5-mini judge at temperature 0). Source: arxiv.org. Saturation forecast: Around February 2028. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 45.1 | #79 |
| 2 | Claude Sonnet 4.6 | 44.2 | #161 |
| 3 | GPT-5.4 | 42.7 | #123 |
| 4 | GPT-5.4 Mini | 37.9 | #389 |
| 5 | GPT-4o | 36.1 | #741 |
| 6 | Claude Haiku 4.5 | 34.7 | #568 |
| 7 | GPT-5.4 Nano | 28.8 | #635 |
| 8 | Phi-4 Multimodal Instruct | 20.3 | #3882 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vsysbench-task-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-13.