PRISM-VLM - Hallucination Resistance: leaderboard
The hallucination-resistance axis of PRISM-VLM (NAVER, 2026): each image comes with a trap question about an absent entity, a misleading premise or an uncertain detail, and the model must refuse or correct it rather than answer. Share of traps handled correctly, in percent, judged by GPT-5; higher is better.
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: Not forecast. 42 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 8B Instruct | 77.9 | #982 |
| 2 | Gemma 4 E4B | 74.6 | #2217 |
| 3 | Claude 3 Haiku | 74.5 | #2629 |
| 4 | Gemini 3 Flash (Minimal) | 73.4 | #159 (Gemini 3 Flash) |
| 5 | Gemini 3.1 Flash Lite (Minimal) | 73.1 | #363 (Gemini 3.1 Flash Lite) |
| 6 | Qwen 3 VL 4B Instruct | 72.9 | #1281 |
| 7 | Gemini 2.0 Flash | 71.6 | #702 |
| 8 | Gemini 2.0 Flash Lite | 71.6 | #1032 |
| 9 | Gemini 2.5 Flash Lite | 71 | #958 |
| 10 | Gemma 4 E2B | 71 | #4124 |
| 11 | GPT-4o Mini | 70.8 | #1627 |
| 12 | Gemini 2.5 Flash (Non-reasoning) | 70.6 | #452 (Gemini 2.5 Flash) |
| 13 | Qwen 3.5 4B (Non-reasoning) | 69.9 | #1180 (Qwen 3.5 4B) |
| 14 | Qwen 3.5 9B (Non-reasoning) | 68 | #803 (Qwen 3.5 9B) |
| 15 | Claude Haiku 4.5 | 67.5 | #568 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=prism-vlm-hallucination-resistance · How It Works · Data refreshed daily, snapshot 2026-11-12.