NAMEBench - Asserted Object Hallucination: leaderboard
NAMEBench (2026), captions of 2,000 randomly selected COCO images per model; hallucinated asserted or existential object mentions, excluding non-assertive mentions. This is the captioning-model outcome, not detector false-positive rate or a disposition score. Hallucination Rate (%); lower is better.
Metric: Hallucination Rate (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 VL 7B | 6.6 | #643 |
| 2 | InternVL3-8B | 7 | #620 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=namebench-asserted-object-hallucination · How It Works · Data refreshed daily, snapshot 2026-10-20.