August 7, 2025
5 min read
Ben Dickson
Google's new diffusion AI agent mimics human drafting and revision to deliver superior enterprise research and complex report generation.
Google’s new diffusion AI agent mimics human writing to improve enterprise research
Google researchers have developed a new AI research framework, Test-Time Diffusion Deep Researcher (TTD-DR), that outperforms leading systems from OpenAI, Perplexity, and others on key benchmarks. Inspired by the human process of drafting, searching for information, and making iterative revisions, TTD-DR uses diffusion mechanisms and evolutionary algorithms to produce comprehensive and accurate research on complex topics. This approach could power bespoke research assistants for enterprises, tackling high-value tasks like competitive analysis or market entry reports—areas where standard retrieval augmented generation (RAG) systems often fall short.The limits of current deep research agents
Deep research (DR) agents aim to address complex queries by using large language models (LLMs) to plan, search, and synthesize detailed reports. They employ test-time scaling techniques such as chain-of-thought (CoT), best-of-N sampling, and Monte-Carlo Tree Search. However, most publicly available DR agents follow rigid linear or parallel workflows that separate planning, searching, and content generation phases. This separation limits interaction and correction between phases, causing loss of global context and missed connections between information pieces.“This indicates a fundamental limitation in current DR agent work and highlights the need for a more cohesive, purpose-built framework for DR agents that imitates or surpasses human research capabilities.”
A new approach inspired by human writing and diffusion
Unlike linear AI agents, human researchers work iteratively: starting with a high-level plan, drafting, and then revising multiple times while searching for new information to strengthen arguments. Google’s researchers modeled this process using a diffusion model augmented with retrieval. Diffusion models, commonly used in image generation, start with a noisy draft and progressively refine it into a detailed output. TTD-DR treats research report creation as a diffusion process, refining an initial “noisy” draft into a polished final report through two core mechanisms:- Denoising with Retrieval: Starting from a preliminary draft, the agent iteratively formulates new search queries, retrieves external information, and integrates it to correct inaccuracies and add detail.
- Self-Evolution: Each component of the agent (planner, question generator, answer synthesizer) independently optimizes its performance using evolutionary algorithms, improving the quality of context for each revision step. This synergy results in reports that are more accurate and logically coherent, with improved fluency and structure suitable for business documents.
- AI agents: Capabilities, risks, and growing role
- How to use Google Gemini for smarter crypto trading
- The future of cryptocurrency explained: what's changing and why it matters
TTD-DR in action
The framework was built and tested using Google’s Agent Development Kit (ADK) with Gemini 2.5 Pro as the core LLM. It was benchmarked against commercial and open-source systems including OpenAI Deep Research, Perplexity Deep Research, Grok DeepSearch, and GPT-Researcher. Evaluations used the DeepConsult benchmark for long-form reports and multi-hop reasoning benchmarks like Humanity’s Last Exam (HLE) and GAIA. TTD-DR consistently outperformed competitors, achieving win rates of 69.1% and 74.5% over OpenAI Deep Research on different datasets, and surpassing it on multi-hop reasoning benchmarks by up to 7.7%.The future of test-time diffusion
While current work focuses on text-based reports using web search, the framework is designed to be adaptable. Future extensions could include generating complex software code, detailed financial models, or multi-stage marketing campaigns—iteratively refined with new information and feedback from specialized tools. Google’s research scientist Rujun Han envisions this draft-centric approach becoming foundational for a wide range of complex, multi-step AI agents.Source attribution: Originally published at VentureBeat on August 6, 2025.