SelfCheckGPT: leaderboard
Black-box hallucination detection benchmark/method using repeated samples from the same model. Inconsistent sampled claims are treated as signals of possible factual fabrication.
Metric: Resistance (100 - SelfCheckGPT). Source: arxiv.org. 9 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | InstructBLIP | 88.32 |
| 2 | LLaVA-v1.5 | 87.43 |
| 3 | FAVOR | 86.22 |
| 4 | Llama-VID | 85.14 |
| 5 | MPLUG-Owl2 | 83.93 |
| 6 | Chat-Univi | 83.2 |
| 7 | Video-LLaVA | 80.03 |
| 8 | Video-Llama | 77.01 |
| 9 | Valley | 74.31 |
Interactive version: theaggregate.ai/benchmark?slug=selfcheckgpt · How It Works · Data refreshed daily, snapshot 2026-09-05.