VSysBench - Misaligned: 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. This board covers the half where the user explicitly tells the model to ignore the system message. Score is the joint satisfaction rate in percent; higher is better.
Metric: Joint satisfaction rate (%; per-sample task score counted only when the judged constraint score reaches 0.8; misaligned half, the user explicitly contradicts the system message with a direct jailbreak prefix; 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 June 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 31.5 | #123 |
| 2 | Claude Opus 4.7 | 29.5 | #79 |
| 3 | Claude Sonnet 4.6 | 25.8 | #161 |
| 4 | GPT-5.4 Mini | 23.2 | #389 |
| 5 | Claude Haiku 4.5 | 17.6 | #568 |
| 6 | GPT-5.4 Nano | 16.1 | #635 |
| 7 | GPT-4o | 11.1 | #741 |
| 8 | Phi-4 Multimodal Instruct | 2.6 | #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-misaligned · How It Works · Data refreshed daily, snapshot 2026-11-13.