Deep researcher with test-time diffusion
13 hours ago
- #AI Research
- #Google Cloud
- #Diffusion Models
- Introduction of Test-Time Diffusion Deep Researcher (TTD-DR), a framework for drafting and revising research reports using high-quality retrieved information.
- TTD-DR models research report writing as a diffusion process, refining a messy first draft into a high-quality final version.
- Two new algorithms introduced: component-wise optimization via self-evolution and report-level refinement via denoising with retrieval.
- TTD-DR achieves state-of-the-art results in long-form report writing and multi-hop reasoning tasks.
- Backbone DR design includes research plan generation, iterative search, and final report generation stages.
- Component-wise self-evolution enhances each stage's agents through initial states, environmental feedback, revision, and cross-over.
- Report-level denoising with retrieval uses search tools to progressively improve the draft report.
- TTD-DR outperforms OpenAI Deep Research across benchmarks, with a 74.5% win rate in long-form research report generation.
- Ablation study shows incremental improvements with each added method, with TTD-DR being more efficient than OpenAI DR.
- TTD-DR is available on Google Cloud Platform via Google Agentspace, implemented with Google Cloud Agent Development Kit.