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

eMoT: evolving Memory-of-Thought via Symbolic Anchoring and Memory Corrosion arXiv:2606.02054v1 Announce Type: new Abstract: 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 reasoning as a transient, one-off generation process rather than retaining an