LLLoCoT Framework Enables Non-Autoregressive Latent Reasoning via Looped Transformers

LLLoCoT replaces left-to-right autoregressive chain-of-thought generation with an iterative refinement process using a compact latent workspace. A shared transformer repeatedly updates latent slots to allow for parallelized latent token sampling and faster reasoning computation.