CaLR Uses Causal Latent Revision to Improve Diffusion-Language-Model Reasoning

Summary

Autoregressive models can become locally greedy, while diffusion language models do not naturally enforce the causal structure needed for reliable reasoning. The paper introduces Causal Latent Revision (CaLR), which formulates reasoning as constrained optimization in latent space. CaLR uses a causal topology matrix derived from an expert model and implicit differentiation to guide revisions of intermediate thoughts. These revisions occur during parallel generation and are intended to maintain logical consistency while enabling dynamic self-correction. The authors report state-of-the-art performance among diffusion language models on complex benchmarks, with results that surpass strong autoregressive baselines. They also report improved robustness on constrained tasks such as Sudoku.