Horangi 4 - Ko-HalluLens (Nonexistent Entities): leaderboard
Ko-HalluLens nonexistent-entity test, Korean questions about entities that do not exist, as run by the Horangi Korean LLM Leaderboard 4 (W&B Korea). Refusal rate (%), higher is better; the remainder are false acceptances.
Metric: Refusal rate (%). Source: horangi.ai. 105 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) (High) | 98.99 | #64 (Gemini 3.1 Pro (Preview)) |
| 2 | Qwen 3.7 Max | 97.96 | #85 |
| 3 | EXAONE 4.5 33B | 97 | #406 |
| 4 | Qwen 3 4B 2507 Instruct | 96 | #768 |
| 5 | GLM-5.2 (High) | 96 | #80 (GLM-5.2) |
| 6 | GPT-5 Mini (2025-08-07) | 95 | #175 |
| 7 | Gemini 3.5 Flash | 95 | #66 |
| 8 | GLM-5-Turbo | 95 | #121 |
| 9 | Qwen 3 235B A22B 2507 (Thinking) | 94.95 | #263 |
| 10 | K-EXAONE (Thinking) | 94 | #387 (K-EXAONE) |
| 11 | EXAONE 3.5 7.8B Instruct | 94 | #856 |
| 12 | Qwen 3.5 397B A17B | 93 | #154 |
| 13 | Qwen 3.7 Plus | 92.93 | #98 |
| 14 | Gemma 4 31B (IT) (Thinking) | 91 | #178 (Gemma 4 31B (IT)) |
| 15 | A.X 4.0 | 91 | #436 |
Interactive version: theaggregate.ai/benchmark?slug=horangi-4-ko-hallulens-nonexistent-entities · How It Works · Data refreshed daily, snapshot 2026-11-05.