Diffusion Language Models Change Speed Across Tasks and Languages
DiffusionGemma needed more than three times as many forward passes per token for story writing as for grammar correction.
Researchers tested two diffusion models across 10 tasks, finding that copying or lightly editing input required less computation than generating new text. Output novelty and word rarity explained much of the gap. Languages requiring more tokens also needed more passes per token, compounding their computational cost. On math problems, weighting candidate answers by decoding speed beat majority voting by up to 4.2 points at no extra cost.
HF Daily Papers' note
Researchers tested two diffusion models across 10 tasks, finding that copying or lightly editing input required less computation than generating new text. Output novelty and word rarity explained much of the gap. Languages requiring more tokens also needed more passes per token, compounding their computational cost. On math problems, weighting candidate answers by decoding speed beat majority voting by up to 4.2 points at no extra cost.
HF Daily Papers' note
score 5