PRISM-VLM - Instruction Following: leaderboard

Metric: Per-axis mean score (%; mean compliance with 1-3 verifiable IFEval-style formatting constraints appended to each item; over PRISM-VLM's 6,238 items recycled from 15 public VLM benchmarks (five seeds of 100 items per benchmark); GPT-5 (low effort) synthesizes the perturbations and grades the open-ended axes; each model at the lowest reasoning effort its provider exposes, temperature 0). Source: arxiv.org. Saturation forecast: Around December 2026. 42 models tracked.

Top models

#ModelScore
1GPT-5 Mini (Minimal)87.5
2GPT-5.4 Mini87.1
3GPT-5.4 Nano86
4GPT-4o Mini84.3
5GPT-5 Nano (Minimal)83.7
6Qwen 3 VL 8B Instruct82.3
7GPT-4.1 Nano81.7
8Gemini 2.5 Flash (Non-reasoning)81.4
9Gemini 3 Flash (Minimal)81
10GPT-4.1 Mini80.5
11Molmo2-8B79.8
12Gemini 2.0 Flash79.1
13Qwen 3 VL 4B Instruct78.9
14Claude 3 Haiku77.8
15Nova 2 Lite77.4

Interactive version: theaggregate.ai/benchmark?slug=prism-vlm-instruction-following · How It Works · Data refreshed daily, snapshot 2026-09-26.