Google introduces Deep Researcher with Test-Time Diffusion, redefining AI research

Test-Time Diffusion Deep Researcher (TTD-DR) is a new framework developed by Google Research for AI-based research agents. Unlike traditional systems that combine tools statically, TTD-DR models research as an iterative process similar to the human method of writing, reviewing, and deepening arguments. The central idea is to treat the first draft as a "noisy" version of the text, which is gradually refined through external searches and fact-checking. This process, inspired by diffusion models, allows the agent to incorporate new facts, correct inconsistencies, and strengthen the report with each cycle. The framework consists of three main stages: generation of a structured research plan, iterative search with information retrieval (RAG), and a self-evolution mechanism that evaluates, revises, and combines multiple responses to achieve higher quality. In tests, TTD-DR outperformed competing agents, including OpenAI Deep Research, achieving a 74.5% win rate on long-form report generation tasks and better performance on multi-step reasoning benchmarks. It also demonstrated greater efficiency, delivering superior results with the same latency. According to Google, TTD-DR represents a significant step toward truly autonomous research agents, capable of acting as scientific assistants, technical analysts, and innovation copilots. A version of the system is already available on Google Agentspace, integrated with the Google Cloud Agent Development Kit.