Learning When Diffusion Language Models Should Regenerate

The 30-second take

• On Sudoku, hidden states carry error information that token confidence does not reliably show, and fixing one token in a frozen context usually reprints the error.
• On five reasoning benchmarks with LLaDA-8B-Base, DiVR raises average accuracy of the fine-tuned generator from 35.3 percent to 43.4 percent.
• Teams already decoding with diffusion language models can add this check at 1.9 to 3.5 times the cost of greedy decoding, below the search baselines at 3.8 to 7.4 times.

What the paper actually did

Remasking lets a diffusion language model revisit text it has already produced, but a correction only helps if the model knows when and what to revise. A diagnosis on Sudoku finds two gaps. Hidden states contain error information that token confidence does not reliably expose. Remasking an isolated error usually reproduces it when the surrounding context stays fixed. Those findings motivate DiVR, a verifier that predicts whether a partially generated sequence will reach a correct answer. It is trained on the final correctness of the model's own trajectories. DiVR decides when to regenerate a block and which alternative to keep. The generator is not modified. On five reasoning benchmarks with LLaDA-8B-Base, DiVR raises the fine-tuned generator's average accuracy from 35.3 percent to 43.4 percent, at 1.9 to 3.5 times the cost of greedy decoding. That is above search baselines that spend 3.8 to 7.4 times greedy cost. The method also complements supervised fine-tuning and reinforcement learning, and it generalizes to other generators, including Dream.

What makes this disruptive

Remasking sounds like self-correction and, on the Sudoku diagnosis, often is not: the local token looks editable, and the fixed neighbors pull the same error back. DiVR does not pretend confidence is a verifier. It learns, from the model's own finished trajectories, whether the current partial sequence is worth continuing. Regeneration is then a block decision, which is a cheaper intervention than a broad search, on the cost multiples the abstract reports.

Why it matters (outside the lab)

Diffusion language models are being tried on reasoning tasks where a single wrong token ruins the chain. A verifier that lifts LLaDA-8B-Base from 35.3 percent to 43.4 percent average accuracy, without retraining the generator, is a module a lab can add beside fine-tuning or reinforcement learning rather than instead of them. The cost band, 1.9 to 3.5 times greedy, is the budget for that add-on.

Limitations & open questions

The accuracy figures are an average over five reasoning benchmarks for one fine-tuned generator, LLaDA-8B-Base. The abstract says DiVR generalizes to other generators, including Dream, and does not give those accuracies. The verifier is trained on the model's own final correctness, so it can share that model's blind spots. The Sudoku study is a diagnosis, not one of the five scored benchmarks. Cost remains above a single greedy pass.