Deep researcher with test-time diffusion

Google Research just published something that quietly reframes how AI agents should write research reports: treat the first draft as noise, and let search results denoise it, the same way diffusion models turn static into images. They're calling it Test-Time Diffusion Deep Researcher, or TTD-DR, and the pitch is that most existing deep-research agents skip the part that actually makes human research good, which is drafting something rough, finding out what's missing, and revising in a loop.

The system starts by generating a structured plan from a user's query, then produces a preliminary draft that acts as scaffolding for everything after. From there it runs an iterative search loop: one sub-agent writes search questions based on the plan and prior answers, another goes and fetches summarized answers from actual sources. Each new answer gets fed back into the draft, either patching gaps or confirming what's already there, and the cycle repeats until the search process wraps and a final agent stitches together the polished report.

Google also bolted on a self-evolution step that works underneath all of this. At each stage, the system generates several candidate answers, has an LLM judge score them on things like helpfulness and comprehensiveness, revises based on that feedback, and then merges the strongest revised variants into one output. It's essentially evolutionary selection applied to intermediate research steps, not just the final report.

On the numbers, TTD-DR beat OpenAI's Deep Research with a 74.5% win rate on long-form report generation using the DeepConsult benchmark, and it edged out OpenAI DR by 7.7% and 1.7% on HLE-Search and GAIA, two benchmarks built around multi-hop reasoning. An ablation study showed the backbone system alone actually underperformed OpenAI DR; it was adding self-evolution and then the denoising-with-retrieval mechanism that pushed results past the competition, with the retrieval step delivering the biggest jump. Google also reports it's more efficient on a latency-versus-quality basis, hitting better win rates at the same response time.

Google built this on Gemini-2.5-pro and has already shipped a version through Google Agentspace using its Agent Development Kit, so this isn't purely a lab exercise. The team credits the gains to what they call a 'draft-first' design, keeping the whole process anchored to an evolving document rather than a pile of disconnected tool calls that lose context along the way.