Causal Shortcut Learning Improves Diffusion Language Model Reasoning

The Causal Shortcut Learning (CSL) framework enhances Diffusion Language Models (DLMs) by using parallel prioritized masking on extracted token chains. This method helps DLMs navigate exponentially large exploration spaces, improving both answer convergence efficiency and reasoning accuracy compared to standard random masking.