We introduce GuideEvolve, a self‑evolving agent framework that combines a task‑specific expert‑evaluation evolutionary pipeline with three autonomous regulatory mechanisms (the 3Ts): Think Fast and Slow, Throwback, and Termination by choice. The expert model scores each generation, steering the search toward task‑optimal solutions, while the 3Ts handle cognitive fast/slow switching, temporal rollback for performance recovery, and adaptive termination based on predefined thresholds. Benchmarks on MGSM and MMLU show that GuideEvolve reaches 89.8% accuracy on MGSM (close to G\"odel Agent’s 90.6%) and 89.1% on MMLU (surpassing 87.9%), reduces variance by 25%‑59%, achieves a 100% evolution success rate versus 82% for the baseline, and cuts required optimization iterations roughly in half (from 30 to 15.1 and 13.4). These results indicate improved robustness and deployment readiness.
Review