Source: Anhui University / EverMind (Shanda Group) / Southeast University — 2026-07-06
Summary
EvoAgentBench is a new benchmark for measuring agent self-evolution through "ability transfer" — whether an agent that learns a reusable procedural skill (a search strategy, a debugging routine, a verification method) in one task can carry it into a new, different task. The benchmark spans four domains (web research, algorithmic reasoning, software engineering, and knowledge work) and grounds abilities in trace-derived "Ability Graphs" rather than free-text descriptions.
Key Takeaways
- Targets a gap in existing agent benchmarks: most measure single-task success rate, not whether learned procedures generalize across tasks.
- Four domains — web research, algorithmic reasoning, software engineering, knowledge work — give reasonably broad coverage for a first version of this benchmark type.
- "Ability Graphs" ground transferable skills in actual execution traces rather than relying on agents self-describing what they learned.
- Directly relevant to the growing interest in self-evolving agents that improve without weight updates.