Unlocking the Working Memory of Large Language Models for Latent Reasoning (arxiv.org)

Researchers have introduced a groundbreaking method called Reasoning in Memory (RiM) that enhances the latent reasoning capabilities of large language models (LLMs). Traditional approaches rely on autoregressive generation, which ties reasoning steps to external communication, limiting efficiency and flexibility. In contrast, RiM employs fixed memory blocks of special tokens that allow LLMs to access working memory directly, facilitating the manipulation of information internally without generating intermediate responses. This innovative process can significantly improve reasoning performance while reducing computational overhead. The significance of RiM lies in its ability to streamline the reasoning process in LLMs by enabling them to operate in a more human-like manner, efficiently utilizing working memory for complex reasoning tasks. The method is implemented through a two-stage curriculum: first, it involves predicting reasoning steps after each memory block, followed by a phase where this supervision is removed, honing the model's ability to refine final answers independently. Experimental results indicate that RiM outperforms or matches existing methods across various language model sizes and families, marking a potential shift in how reasoning can be approached in AI systems. This advancement could lead to more sophisticated AI applications that require nuanced reasoning capabilities.