Deep researcher with test-time diffusion

Google Research · 1 year ago · 16

Google introduced Test-Time Diffusion Deep Researcher (TTD-DR), an AI agent that models research report writing as an iterative diffusion process, gradually refining a preliminary draft into a polished final version using retrieval and self-evolution algorithms. TTD-DR achieved a 74.5% win rate compared to OpenAI Deep Research on long-form report generation tasks and outperformed competing systems by 7.7% and 1.7% on multi-hop reasoning benchmarks. The system is now available on Google Cloud Platform through Google Agentspace, enabling more coherent and efficient research agent workflows that better emulate human research patterns.