DiffResearch: Accelerating Deep Research with Diffusion Language Models
Swapping in a diffusion model cut end-to-end research time roughly sixfold across 100 benchmark tasks.
The researchers held the agent framework, retrieval code, prompts and evaluation fixed while comparing diffusion and autoregressive backbones. Planning and writing saw order-of-magnitude reductions in latency. Their dFast mode extracts subqueries before generation finishes, nearly halving median decomposition time without reducing report quality and earning the highest overall RACE score of the three configurations. DiffResearch is released under the Apache-2.0 license.
HF Daily Papers' note
The researchers held the agent framework, retrieval code, prompts and evaluation fixed while comparing diffusion and autoregressive backbones. Planning and writing saw order-of-magnitude reductions in latency. Their dFast mode extracts subqueries before generation finishes, nearly halving median decomposition time without reducing report quality and earning the highest overall RACE score of the three configurations. DiffResearch is released under the Apache-2.0 license.
HF Daily Papers' note
score 5