PhantomBench - Terms (Place): leaderboard
PhantomBench (2026) asks language models about generated terms and entities that do not exist, verified absent from a 2.3-trillion-token corpus, and counts answers that invent content instead of abstaining, as labeled by a Gemini 2.5 Flash judge. This board is the hallucination rate in percent on 1,060 non-existent terms modeled on medical abbreviations, science glossaries and legal terms, when the prompt asks where it was discovered or took place, presupposing it exists; lower is better.
Metric: Hallucination rate (%): share of answers that do not abstain on the Phantom-T subset (1,060 generated non-existent terms modeled on medical abbreviations, science glossaries and legal terms), verified absent from the 2.3-trillion-token Dolma corpus, when the prompt asks where the concept was discovered or took place (presupposing it exists); a Gemini 2.5 Flash judge labels abstention; reasoning-model rates count only generations that reached a final answer; lower is better. Source: arxiv.org. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 8B | 3.21 | #1830 |
| 2 | Qwen 3 14B | 3.3 | #1233 |
| 3 | Llama 3.1 8B Instruct | 3.58 | #4246 |
| 4 | Qwen 3 4B | 6.23 | #2281 |
| 5 | Qwen 2.5 7B Instruct | 6.51 | #2564 |
| 6 | Qwen 3 1.7B | 9.15 | #5520 |
| 7 | Gemma 2 9B (IT) | 12.83 | #2262 |
| 8 | Qwen 3 32B | 14.06 | #960 |
| 9 | Gemini 2.5 Pro | 15.75 | #266 |
| 10 | Gemini 2.5 Flash | 19.34 | #453 |
| 11 | Mistral 7B Instruct (v0.3) | 43.68 | #6653 |
| 12 | DeepSeek R1 Distill Qwen 32B | 58.97 | #1788 |
| 13 | GPT-OSS-20B (Low) | 72.68 | #1147 (GPT-OSS-20B) |
| 14 | Gemma 3 12B (IT) | 85.28 | #1837 |
Interactive version: theaggregate.ai/benchmark?slug=phantombench-terms-place · How It Works · Data refreshed daily, snapshot 2026-11-13.