Continual Learning Exam (Trial 1)
Agents · 2026-10-25
120 expert-authored professional tasks across 12 occupational domains, each a five-level ladder (600 levels) whose later levels build on experience from earlier ones. The agent runs the whole benchmark as one serial stream and must pass a level to unlock the next; a DeepSeek V4 Flash evaluation agent grades each submission against human-written checklists. Trial 1: one submission per level, no feedback. The score is the share of the 600 levels cleared, in percent.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 42.2 |
| GPT-5.6 Sol | 27.8 |
| GPT-5.6 Terra | 27.8 |
| Sonnet 5 | 27.0 |
| Qwen3.8-Max | 25.7 |
| Gemini 3.7 Flash | 25.3 |
| GLM-5.3 | 22.5 |
| DeepSeek V4 Pro 0813 | 21.8 |
| Gemini 3.8 Flash | 21.2 |
| Gemini 3.6 Flash | 21.0 |
| GPT-5.6 Luna | 20.3 |
| Hy3 | 20.3 |