What Makes a Good Unmasking Policy? A Controlled Study of Learned Unmasking in Diffusion Language Models
Open paper
Benchmark
Model
Rank
Results
arithmetic-reasoning-on-gsm8k
OnTrack (Ours)
–
Accuracy: 78.8
code-generation-on-humaneval
OnTrack (Ours)
–
Pass@1: 43.9
Rank counts only results with a code link.