RESEARCH · RESEARCH · #2475
DiffResearch paper releases open diffusion-LM research framework claiming ~6× end-to-end speedup
The authors introduce DiffResearch, an open multi-agent deep-research framework that uses a diffusion language model (dLLM) as the backbone and report that swapping an autoregressive backbone for a diffusion one yields roughly 6× faster end-to-end latency on 100 Deep Research Bench tasks while keeping the agent scaffold, retrieval, prompts, and evaluation fixed. They also present dFast, a decomposition mode that reads intermediate denoising states to extract subqueries (reporting a 1.99× median speedup for decomposition) and say an instance built on a stronger diffusion model ranks above several commercial systems; the code is released under an Apache-2.0 license (arXiv:2610.28909v1).
KEY POINTS
- The authors introduce DiffResearch, an open multi-agent deep-research framework that uses a diffusion language model (dLLM) as the backbone and report that swapping an autoregressive backbone for a diffusion one yields roughly 6× faster end-to-end latency on 100 Deep Research Bench tasks while keeping the agent scaffold, retrieval, prompts, and evaluation fixed.
- They also present dFast, a decomposition mode that reads intermediate denoising states to extract subqueries (reporting a 1.99× median speedup for decomposition) and say an instance built on a stronger diffusion model ranks above several commercial systems; the code is released under an Apache-2.0 license (arXiv:2610.28909v1).
- If validated, diffusion-based LLM backbones could materially reduce latency in multi-step research and multi-agent pipelines and the open Apache-2.0 release enables replication and further evaluation.
WHY IT MATTERS
If validated, diffusion-based LLM backbones could materially reduce latency in multi-step research and multi-agent pipelines and the open Apache-2.0 release enables replication and further evaluation.