ReaLM: Reflection-Enhanced Autonomous Reasoning with Small Language Models