AuthorityBench - Fabricated Citations on True Claims: leaderboard
AuthorityBench (2026) isolates how citation-based authority cues change LLM behaviour by crossing true and false claims with real and fabricated citations in four domains. This board is the hallucination rate in percent on true claims presented with a fabricated citation, judged by Qwen3-8B; lower is better.
Metric: Hallucination rate (%) on true claims presented with a fabricated citation (TC x FC condition): share of non-refused responses a Qwen3-8B judge labels hallucinated (the model denies or distorts the true claim); 15K stratified subset for four models, full dataset for Gemma 3 4B, Llama 3.1 8B Instruct and Phi-4 Mini Instruct; lower is better. Source: arxiv.org. Saturation forecast: Around March 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.2 | 16.28 | #406 |
| 2 | Gemma 4 31B | 22.73 | #394 |
| 3 | Gemma 3 4B | 24.14 | #5064 |
| 4 | GPT-5.4 Mini | 37.27 | #389 |
| 5 | Phi-4 Mini Instruct | 40.63 | #6419 |
| 6 | Llama 3.1 8B Instruct | 43.76 | #4246 |
| 7 | Claude Haiku 4.5 | 47.11 | #568 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=authoritybench-fabricated-citations-on-true-claims · How It Works · Data refreshed daily, snapshot 2026-11-13.