CISE Framework for Reliable Self-Evolution via Conformal Inference

CISE uses conditional conformal inference and online density-ratio estimation to create statistically calibrated reward intervals for LLM self-evolution. This prevents low-cost, imperfect proxy rewards from introducing false positives into the evolutionary feedback loop.