Continual Learning Bench Average Gain
Agents · 2026-05-04
CL-Bench normalizes each task's reward gain over the same system's stateless baseline against its published reference ceiling, then averages the six normalized gains equally.
Top models (higher is better)
| Model | Score |
|---|---|
| Sonnet 4.6 | 0.2 |
| GPT-5.4 | 0.2 |
| Opus 4.7 | 0.2 |
| Gemini 3 Flash Preview | 0.2 |
| Gemini 3.1 Pro Preview | 0.1 |