OmniHalluBench
A 3,540-example benchmark for evaluating omni-modal hallucination in LLMs across text, image, audio, and video modalities.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info omnihallubench
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
2 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Qwen2.5-Omni-7B | 65.1 | omniconfess | Independent | medium | Oct 2, 2026 | arxiv.org |
| Qwen3-Omni-30B-A3B-Thinking | 57.0 | sc | Independent | medium | Oct 2, 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
OmniHalluBench comprises 3,540 hard examples derived from six public datasets, covering judgment and free-form generation tasks. It spans text, image, audio, and video modalities to assess cross-modal conflicts and text-grounded hallucinations.
The benchmark evaluates model performance using category-mean F1 scores. Experiments demonstrate that the proposed OmniConfess method outperforms strong baselines across all included datasets on the primary backbone.