Autoregressive (AR) models tend to be locally greedy, which hampers long‑range coherence, while diffusion language models (DLMs) enable parallel generation but often miss a strict causal structure needed for reliable reasoning. To combine their strengths and address their weaknesses, we introduce Causal Latent Revision (CaLR), a framework that reformulates reasoning as a constrained latent‑space optimization problem. CaLR extracts a causal topology matrix (CTM) from an expert model, applies implicit differentiation to obtain gradients in the latent space, and performs gradient‑guided "thought revision". This dynamic self‑correction adjusts intermediate steps during parallel generation, enforcing logical consistency throughout the output. Empirical results show CaLR achieves state‑of‑the‑art DLM performance on challenging benchmarks, surpasses strong AR baselines, and demonstrates superior robustness on constrained tasks such as Sudoku.
Review