NOLLI - Average: leaderboard
NOLLI (arXiv 2608.04397v1, Table 2): procedurally generated, uniquely solvable English-Korean logic puzzles for 15 models, accuracy pooled over three difficulty tiers calibrated on a reference model. Macro average over all 25 tasks (15 puzzle types across languages); percent, higher is better.
Metric: Accuracy (%; 25-task macro average). Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (Medium) | 84.9 | #33 (GPT-5.5) |
| 2 | Claude Opus 4.8 (Medium) | 78.3 | #39 (Claude Opus 4.8) |
| 3 | Gemini 3.1 Pro (Preview) (Medium) | 73.6 | #64 (Gemini 3.1 Pro (Preview)) |
| 4 | Qwen 3.5 397B A17B (Non-reasoning) | 62.5 | #154 (Qwen 3.5 397B A17B) |
| 5 | Gemma 4 31B (IT) (Thinking) | 48.3 | #178 (Gemma 4 31B (IT)) |
| 6 | GPT-OSS-120B (Medium) | 47.7 | #346 (GPT-OSS-120B) |
| 7 | DeepSeek V4 Flash | 45.8 | #155 |
| 8 | Qwen 3.5 27B | 44.4 | #192 |
| 9 | Qwen 3.5 9B (Non-reasoning) | 32.5 | #391 (Qwen 3.5 9B) |
| 10 | EXAONE 4.0 32B (Thinking) | 25.1 | #615 (EXAONE 4.0 32B) |
| 11 | Llama 4 Maverick | 14.4 | #486 |
| 12 | Solar Open 100B | 14.1 | #415 |
| 13 | Llama 3.1 8B Instruct | 1.5 | #1017 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=nolli-average · How It Works · Data refreshed daily, snapshot 2026-11-05.