CiteAudit (Real-World): leaderboard
Metric: F1 (fraction times 100) for flagging hallucinated references, the hallucinated class positive, on the CiteAudit real-world test set (3,356 references from scholarly manuscripts: 2,889 authentic and 467 naturally occurring hallucinated citations); each language model classifies every reference as real or hallucinated through its API (tool access is not described); higher is better. Source: arxiv.org. Saturation forecast: Around March 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 57.1 | #77 |
| 2 | Claude Sonnet 4.5 | 44.4 | #138 |
| 3 | GPT-5.2 | 33.1 | #105 |
| 4 | Llama 3.3 70B Instruct | 19.4 | #520 |
| 5 | Mixtral 8x7B Instruct (v0.1) | 14.4 | #1049 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=citeaudit-real-world · How It Works · Data refreshed daily, snapshot 2026-10-11.