GDPx: Recursive Self-Improvement
Code · 2026-11-01
GDPx's Recursive Self-Improvement domain: 10 tasks from Extrapolate on ML research engineering: training, evaluating and improving models end to end. Scores are pass@1 over three attempts per task. A component of the GDPx score, which weights its 11 domains equally.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 24.9 |
| Claude Sonnet 5.5 | 22.9 |
| GPT-6.1 Sol | 22.8 |
| Gemini 4 Carbon | 19.1 |
| Claude Opus 5.5 | 17.6 |
| GPT-6.1 Terra | 17.4 |
| Claude Fable 5.1 | 16.5 |
| Muse Wonder Full | 15.8 |
| Gemini 4 Argon | 14.1 |
| Claude Haiku 5.5 | 11.3 |
| GPT-6 Luna | 7.8 |
| Kimi K3 | 7.5 |
| Muse Wonder | 7.3 |
| GLM-5.3 | 7.1 |
| MiniMax M3.1 Flash | 6.6 |
| Grok 4.7 | 5.2 |
| Muse Spark Open | 1.8 |