The Quality Ceiling

NLD matches Qwen3-8B's AR accuracy. That's the headline. But "matches" is doing important work here.

In the paper's evaluation, NLD's diffusion decoding achieves 70.28% accuracy on their benchmark suite — essentially tying Qwen3-8B's AR baseline. This is the first time a diffusion LM has closed the gap with pure AR generation. It's a real achievement.

But "matching" is not "exceeding." The most accurate version of NLD is still AR mode on the same weights. Diffusion mode trades a small amount of quality for a large amount of speed. Self-speculation mode preserves accuracy (the AR verification is lossless) but adds the system complexity of managing draft-verify cycles.

The deeper question: for what fraction of real-world tasks does the small quality gap matter? If diffusion mode loses 0.3% on MMLU but generates 3x faster, most users will never notice the difference. But for tasks requiring exact reasoning — mathematical proofs, legal analysis, code generation for safety-critical systems — even small quality differences matter.

The paper doesn't answer this question. It reports benchmark averages, not per-task breakdowns. The real quality ceiling remains an open question for each deployment.