"Add a quality knob that does not lengthen output Traditional reasoning models improve answer quality by producing longer and longer thinking traces, which costs memory and time proportional to length. Diffusion language models offer an orthogonal lever: 'you can kind of like control the number of denoising steps... the model is actually able to do error correction, it's able to improve its own answer without having to make it longer and longer, which saves memory.' When designing any iterative process, ask whether quality improvements must increase output size or whether in-place refinement is possible — the latter is almost always more efficient. Source: Arash Vahdat, Latent Space Podcast, Diffusion LLMs with Inception AI"