Thematic Generalization V2
General QA · 2026-03-16
Thematic Generalization V2 Top-1 Accuracy on the pinned 703-case hard subset tests whether models can infer a latent theme from positive examples and anti-examples and identify the one true match among close distractors. Higher is better; inverse-rank results use the separate thematic-generalization-v2-inverse-rank-score slug.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.1 Pro Preview | 91.5 |
| GPT-5.4 | 90.9 |
| Opus 4.6 | 90.0 |
| Sonnet 4.6 | 88.5 |
| Opus 4.7 | 86.8 |
| GLM-5.1 | 85.9 |
| Kimi K2.5 | 84.8 |
| Qwen3.5 397B A17B | 82.4 |
| Gemini 3.1 Flash-Lite Preview | 82.1 |
| DeepSeek-V3.2 | 81.8 |
| Grok 4.20 | 81.5 |
| Qwen3.6 Plus (2026-04-02) | 81.5 |
| GPT-5.4 Mini | 80.8 |
| Doubao Seed 2.0 Pro | 77.0 |
| Qwen3.5 122B A10B | 76.5 |
| Gemma 4 31B IT | 76.1 |
| Qwen3.5-27B | 71.3 |
| MiMo-V2-Pro | 68.8 |
| Trinity Large Thinking | 66.1 |
| ERNIE 5.0 | 65.3 |
| MiniMax M2.7 | 63.4 |
| Mistral Large 3 675B Instruct 2512 | 34.0 |
| Mistral Medium 3.1 | 32.3 |