CiteCheck (Few-Shot): leaderboard
Citation hallucination detection from the CiteCheck paper (Khajavi et al., 2026) with 27 labelled validation examples in the prompt and no search: an LLM labels each of 792 held-out physics citations as exact, minor hallucination or major hallucination. Macro-F1 over the three classes in percent; higher is better.
Metric: Macro-F1 (%; three-way classification of a physics citation as exact, minor hallucination (corrupted metadata) or major hallucination (fabricated), 792-citation held-out test split; LLM prompted directly with the shared rubric, temperature 0, 27 labelled validation examples, no web search). Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 73.7 | #158 |
| 2 | GPT-5.4 | 67.3 | #121 |
| 3 | Gemini 2.5 Flash | 62.2 | #452 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=citecheck-few-shot · How It Works · Data refreshed daily, snapshot 2026-11-12.