Search papers, labs, and topics across Lattice.
The paper introduces Reasoning in Memory (RiM), a novel latent reasoning method that replaces autoregressive generation of reasoning steps with fixed-sequence memory blocks to unlock LLMs' working memory. RiM uses a two-stage curriculum, first grounding memory blocks by predicting explicit reasoning steps, then iteratively refining the final answer after each block without step-level supervision. Experiments on reasoning benchmarks demonstrate that RiM matches or exceeds existing latent reasoning methods while being more compute-efficient by avoiding autoregressive generation.
LLMs can match or exceed the reasoning performance of autoregressive methods, but with significantly improved compute efficiency, by using fixed-sequence "memory blocks" for latent reasoning.
To improve the reasoning capabilities of large language models, test-time compute is typically scaled by generating intermediate tokens before the final answer. However, this couples reasoning to autoregressive generation and thereby conflates internal computation with external communication. In contrast, human cognition can use working memory to hold and manipulate information internally without the need to externalize intermediate thoughts. Drawing on this principle, we introduce Reasoning in Memory (RiM), a latent reasoning method that replaces the autoregressive generation of reasoning steps with memory blocks. These memory blocks are fixed sequences of special tokens that unlock the working-memory capacity of large language models. Since they are fixed rather than generated, they can be processed in a single forward pass, enabling compute-efficient latent reasoning. To operationalize these memory blocks, we employ a two-stage curriculum. First, we ground them by predicting explicit reasoning steps after each memory block. Second, we discard this step-level supervision and iteratively refine the final answer after each memory block. Our experiments on reasoning benchmarks show that, across language models of different families and sizes, RiM matches or exceeds existing latent reasoning methods while avoiding the autoregressive generation of thoughts. These results demonstrate that large language models can be trained to use working memory as an effective mechanism for latent reasoning.
