CL-Bench, a new benchmark testing whether AI agents actually improve with experience across sequential real-world tasks, found that even the best system -- Claude Sonnet 4.6 using full-context in-context learning -- achieved only a 25.4% normalized gain over its own stateless baseline

Researchers from UC Berkeley and Snorkel AI introduced CL-Bench, a new benchmark to evaluate how well AI systems improve with sequential experience in real-world tasks. They measured the "gain metric" to isolate learning from prior capabilities. The best performing system, Claude Sonnet 4.6 using full-context in-context learning, achieved a 25.4% normalized gain over its own baseline.