Test-Time Scaling for Diffusion Language Models with Frequency-Aware Remasking

A research paper proposes frequency-aware remasking, a test-time scaling method for diffusion language models (DLMs) that replaces instantaneous confidence or entropy signals with frequency-based ones when deciding which tokens to commit or remask during iterative denoising. The authors argue existing certainty-aware strategies can be misleading on difficult reasoning tasks because their local signals are unreliable. The work targets output quality in DLM text generation without requiring model retraining.

Diffusion language models DLMs generate text through iterative denoising, where decisions about which tokens to commit or remask strongly affect output quality. Existing certainty-aware strategies rely on instantaneous confidence or entropy, but these local signals can be misleading on difficult r