j-cot-chain-of-thought-in-j-space-cc72b48d·1 events·first seen Aliases: J-CoT: Chain-of-Thought in J-Space
J-CoT introduces a recurrent reasoning framework that operates in 'J-space', a vocabulary-indexed coordinate system within a model's hidden representations, allowing intermediate reasoning states to be carried forward without requiring full verbalization or dense hidden-state recurrence. At each cycle boundary, the model expresses its intermediate state as vocabulary-indexed coefficients (a 'J-thought'), which are then mapped back into the hidden representation for the next step. J-CoT-Zero matches or exceeds the strongest latent-reasoning baseline on every evaluated benchmark, and J-CoT-Train achieves top scores across mathematical, scientific, coding, and structured path-reasoning tasks. The approach occupies a novel middle ground between natural-language chain-of-thought and fully continuous latent reasoning methods.