Aspire finds self-evolution still loses its gains

Aspire tested self-evolving agents against 520 hidden, expert-authored items across six goals, and found current systems often finish training or harness-editing loops without retaining above-base gains. The Aspire benchmark paper says the strongest evolved harness still trails an engineered Qwen-Agent reference and that local self-evaluations fail to transfer to hidden tasks. Treat self-training as an experiment loop that needs sealed external evaluations, not as a capability upgrade; OpenClaw’s 16,000-PR release shows why operating harness changes also need controlled rollback.