Reduce the computational overhead of quantum-inspired remasking

Reduce the computational overhead of the quantum-inspired remasking procedure, which requires additional model evaluations to estimate remasking rates.

Background

The quantum-inspired remasking (QIR) procedure estimates a remasking rate by evaluating the diffusion LLM on modified sequences in which candidate unmasked tokens are replaced by mask tokens, then taking reciprocal score estimates. Because these evaluations are performed for candidate positions during generation, QIR incurs additional inference cost compared with ordinary sampling.

The paper identifies reducing this computational overhead as an explicit open problem. The implementation samples only a limited number of candidate positions to control cost, but this is an approximation rather than a resolution of the underlying efficiency issue.