Diffusion language models can rewrite their own drafts — but rewriting the wrong span just reproduces the same mistake.
Remasking lets these models revisit generated text, yet they have never had a good sense of when a revision is worth it. DiVR adds a verifier that predicts whether a partial sequence will still reach the right answer, then decides when to regenerate a block and which alternative to keep.
On LLaDA-8B-Base across five reasoning benchmarks, average accuracy rose from 35.3% to 43.4% at 1.9–3.5× the cost of greedy decoding — while search baselines needed 3.8–7.4× the compute for worse results.
The practical payoff: a real accuracy lift without touching or retraining the generator, and the verifier even transfers to other generators like Dream. The catch is still multiplied compute, so latency-sensitive deployments need to do the math first.
Would you trade 2–3× compute for an 8-point accuracy gain on your own reasoning tasks?
#
DiffusionLM# #
LLMReasoning#