What does FFN compression change downstream? Same-state causal restoration in diffusion language models

Diffusion language models (DLMs) enable flexible, parallel generation, but their iterative denoising remains computationally expensive, motivating increasingly aggressive compression.