Chain-of-Thought (CoT) reasoning often boosts language‑model performance by allocating extra computation before answering. Traditional CoT expresses this as an autoregressive token sequence, while latent reasoning replaces those tokens with compact continuous states; however, most approaches still enforce a left‑to‑right dependency among latent vectors. LLoCoT (Looped Latent CoT) replaces left‑to‑right latent generation with iterative refinement of a compact latent workspace. A shared transformer is reapplied for a few refinement iterations, jointly updating latent slots based on the prompt and the evolving workspace state. After refinement, a probabilistic head predicts a distribution, from which latent tokens are sampled in parallel and used to condition an autoregressive decoder for answer generation. Training leverages continuous representations derived from explicit CoT together with a final‑answer prediction loss and likelihood‑based supervision of the latent states. Across HumanEval and MBPP, LLoCoT attains the highest mean among evaluated methods, matching the accuracy of the explicit‑CoT baseline Reasoning SFT while outperforming the base model, answer‑only SFT, and NF‑CoT. Compared to Reasoning SFT, LLoCoT reduces time to the first answer token by roughly $36\times$, cuts reasoning‑phase latency by about $42\times$, and increases end‑to‑end throughput by $9.2\%$. This design replaces serial thought generation with parallel latent‑slot refinement while retaining probabilistic latent modeling and autoregressive answer decoding.
Review: LLoCoT demonstrates that looped refinement can achieve substantial speed‑ups without sacrificing reasoning quality, offering a promising direction for efficient large‑model inference.