PRISM-VLM - Hallucination Resistance: leaderboard

Metric: Per-axis mean score (%; share of trap questions (absent entity, misleading premise, uncertain image) refused or corrected; 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 March 2028. 42 models tracked.

Top models

#ModelScore
1Qwen 3 VL 8B Instruct77.9
2Gemma 4 E4B74.6
3Claude 3 Haiku74.5
4Gemini 3 Flash (Minimal)73.4
5Qwen 3 VL 4B Instruct72.9
6Gemini 2.0 Flash71.6
7Gemini 2.0 Flash Lite71.6
8Gemini 2.5 Flash Lite71
9Gemma 4 E2B71
10GPT-4o Mini70.8
11Gemini 2.5 Flash (Non-reasoning)70.6
12Qwen 3.5 4B (Non-reasoning)69.9
13Qwen 3.5 9B (Non-reasoning)68
14Claude Haiku 4.567.5
15Nova 2 Lite67.1

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