J-CoT: A New Recurrent Reasoning Framework for Language Models

A recent research article presents J-CoT, a recurrent reasoning framework functioning within J-space, a coordinate system that indexes vocabulary in the hidden representations of a language model. While chain-of-thought prompting enhances reasoning by maintaining intermediate states throughout various steps, relying solely on natural language as a recurrent interface proves limiting, as many temporary calculations do not necessitate complete verbalization. Current latent-reasoning techniques transmit continuous hidden states but lack clear methods for choosing and structuring the information required for subsequent steps. J-CoT resolves this by offering an intermediate interface that is linguistically grounded yet does not need a decoded sentence. The model performs computations in its entire hidden space during each cycle, and at the cycle's end, J-CoT organizes information using the J-space coordinate system. The paper can be found on arXiv under ID 2607.21981.