AI TrackerFrontier desk · AEST
Back to news
normalaitech

Learning When Diffusion Language Models Should Regenerate

arXiv cs.CL · 13 Nov 2026, 04:58 AEST

Remasking allows diffusion language models to revisit generated text, but effective correction requires knowing when and what to revise. DiVR, a verifier, predicts whether a partially generated sequence will reach a correct answer and improves accuracy significantly.

Key facts

Read aloud

Remasking allows diffusion language models to revisit generated text, but effective correction requires knowing when and what to revise. DiVR, a verifier, predicts whether a partially generated sequence will reach a correct answer and improves accuracy significantly. Our diagnosis on Sudoku reveals two gaps. Hidden states contain error information that token confidence does not reliably expose, and remasking an isolated error usually reproduces it when the surrounding context remains fixed. Trained on the final correctness of the model's own trajectories, DiVR decides when to regenerate a block and which alternative to retain, while the generator itself stays unchanged. On five reasoning benchmarks with LLaDA-8B-Base, DiVR raises the average accuracy of the fine-tuned generator from thirty five point three percent to forty three point four percent at one point nine dollars$-three point five dollars imes$ the cost of greedy decoding, above search baselines that use three point eight dollars$-seven point four dollars imes$.

Sources in this cluster