Deep Researcher with Test-Time Diffusion
Rujun Han, Yanfei Chen, Zoey CuiZhu, Lesly Miculicich, Guan Sun, Yuanjun Bi, Weiming Wen, Hui Wan, Chunfeng Wen, Solène Maître, George Lee, Vishy Tirumalashetty, Emily Xue, Zizhao Zhang, Salem Haykal, Burak Gokturk, Tomas Pfister, Chen-Yu Lee
2025-07-28
Summary
This paper talks about Deep Researcher with Test-Time Diffusion (TTD-DR), an AI system that writes high-quality research reports by starting with a rough draft and improving it step-by-step using information from the internet and self-improving techniques.
What's the problem?
Current AI research tools create reports by splitting tasks into isolated steps, which often leads to missing connections, losing important details, and inefficient information search, making the generated reports less coherent and useful.
What's the solution?
The researchers designed TTD-DR to mimic how humans write by continuously refining one draft through multiple rounds of searching, adding new information, and revising. The system also improves each part of the process by generating multiple options and picking the best, which resembles a peer review.
Why it matters?
This matters because TTD-DR makes AI-generated research more accurate, coherent, and context-aware, enabling better long-form report writing for complex tasks like scientific research, financial analysis, and market intelligence.
Abstract
TTD-DR, a diffusion-based framework, generates high-quality research reports by iteratively refining a preliminary draft with external information and self-evolutionary algorithms, outperforming existing deep research agents.