Benchmarks

← All benchmarks

Continual Learning Bench (CL-Bench)

Agents · 2026-05-04

Continual Learning Bench evaluates stateful agents that adapt across 301 sequential instances in six expert-validated task domains. This headline metric normalizes each task's reward against a reference ceiling and the GPT-5.4 stateless baseline, then averages the six normalized task scores.

Top models (higher is better)