EvoAgentBench: A Benchmark That Isolates Whether Agents Can Turn Past Sessions Into Reusable Procedures

arXiv·medium signal

EvoAgentBench (arXiv:2607.05202, July 6, 2026) isolates procedural self-evolution — whether an agent can turn past sessions into reusable search, debug, and verify procedures — across web research, algorithmic reasoning, software engineering, and knowledge work, under a Train/Extract/Test protocol with held-out non-duplicate tasks (dataset released on HuggingFace). It argues current agent benchmarks test single-episode solving and memory benchmarks test fact retention, missing procedural reuse. For builders: a standardized way to check whether your memory or skill-extraction layer actually transfers know-how to new tasks rather than just storing facts.