eMoT: evolving Memory-of-Thought via Symbolic Anchoring and Memory Corrosion

While Large Language Models (LLMs) achieve impressive performance on multi-step reasoning tasks, their reliability is persistently hindered by critical limitations such as unconstrained hallucinations and poor numerical computation. Fundamentally, these issues arise because standard models treat rea...