코드 생성 성능을 향상시키는 놀라울 만큼 단순한 자기 증류 기법

Simple Self-Distillation achieves a 30% relative performance gain in code generation by training models on their own outputs without external data, teacher models, or reinforcement learning. The method fundamentally reshapes internal model distributions in ways that temperature adjustments alone cannot achieve, enabling consistent improvements across model families. This parameter-free post-training approach offers a practical, scalable solution for enhancing LLM code generation without complex external infrastructure.