Research result · Sep 3, 2026
EvoMap Team Unveils Self-Evolving Agent System Realizing Recursive Self-Improvement via Experience Flow
The EvoMap team has released a self-evolving agent system built from three components: the EvoX agent, the Evolver evolution engine, and the EvoMap experience-flow network. Through the GEP protocol, 357,000 agents share 4.81 million experience assets. On benchmarks including LongWoF-Bench, the team reports gains of 8.7–15.5 percentage points over a standard Skill baseline. The system continues to evolve without updating model parameters.
Original sources (Chinese)
RSI离我们有多远？这家中国团队给出了第一梯队的工程解