Benchmark RadarAI BENCHMARK PROFILE
EvoAgentBench
EvoAgentBench evaluates agent self-evolution via ability transfer across web research, algorithmic reasoning, software engineering, and knowledge work, using trace-grounded abilities and domain graphs with train/test splits.
- Released
- 2026-07-06
- Readiness
- Inspectable
- Primary field
- General AI
Why it matters
Current evaluations don't isolate procedural reuse. This benchmark enables fine-grained diagnosis of experience encoding, routing, and uptake in agent self-evolution.
Motivation
Agent self-evolution in long-horizon LLM systems is largely procedural: useful experience is not merely stored information, but reusable procedures for searching, debugging, and verification.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.