NAME-Bench
Non-Assertive Mention Evaluation Benchmark
A benchmark of 25,497 object mentions from VLM captions on COCO images, labeled as concrete, hallucinated, or non-assertive to evaluate hallucination detectors.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info name-bench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
5 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| LLaVA-1.5-7B | 28.2 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Qwen2.5-VL-3B | 8.9 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| InternVL3-8B | 7.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Qwen2.5-VL-7B | 6.6 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| LLaVA-OneVision-2-8B | 5.7 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
NAME-Bench comprises 25,497 object mentions generated by five VLMs for COCO images. Each mention is annotated for usage (asserted vs. non-assertive) and presence, with non-assertive types further classified into depictive, purpose, negation, and hedged subtypes.
It evaluates twelve hallucination detectors, revealing that they frequently flag non-assertive mentions as hallucinations. The benchmark demonstrates that training with non-assertive negatives reduces false positives, though effectiveness varies significantly across different VLMs.