codelion
Abstract
The Test-Time Diffusion Deep Researcher (TTD-DR) framework uses a diffusion process with iterative refinement and external information retrieval to generate high-quality research reports, outperforming existing methods.
Deep research agents, powered by Large Language Models (LLMs), are rapidly advancing; yet, their performance often plateaus when generating complex, long-form research reports using generic test-time scaling algorithms. Drawing inspiration from the iterative nature of human research, which involves cycles of searching, reasoning, and revision, we propose the Test-Time Diffusion Deep Researcher (TTD-DR). This novel framework conceptualizes research report generation as a diffusion process. TTD-DR initiates this process with a preliminary draft, an updatable skeleton that serves as an evolving foundation to guide the research direction. The draft is then iteratively refined through a "denoising" process, which is dynamically informed by a retrieval mechanism that incorporates external information at each step. The core process is further enhanced by a self-evolutionary algorithm applied to each component of the agentic workflow, ensuring the generation of high-quality context for the diffusion process. This draft-centric design makes the report writing process more timely and coherent while reducing information loss during the iterative search process. We demonstrate that our TTD-DR achieves state-of-the-art results on a wide array of benchmarks that require intensive search and multi-hop reasoning, significantly outperforming existing deep research agents.
Community
🎉 We've implemented TTD-DR in OptiLLM!
Really impressed by this approach - the diffusion-inspired iterative refinement for research generation is brilliant. We've successfully implemented the full algorithm and tested it extensively.
Implementation highlights:
• Full gap analysis and iterative denoising pipeline
• Integrated web search with Selenium for real-time information gathering
• Smart reference aggregation from multiple sources
• Works with any OpenAI-compatible model (GPT-4, Claude, Llama, Mistral, etc.)
Results from our testing:
• Evaluated on 47 complex research queries
• Generated reports with 15-30+ real web sources each
• Topics ranged from investment analysis to emerging tech landscapes
• Quality genuinely rivals human research analysts
What we learned:
• The iterative refinement really makes a difference - each denoising step noticeably improves quality
• Gap analysis is crucial for identifying what's missing in the initial draft
• Smaller models (7B) can produce excellent research when augmented with TTD-DR
• The approach effectively eliminates hallucination on current events
You can try it out:
pip install optillm
# Then use "deep_research-<your-model>" as the model name
Code: https://github.com/codelion/optillm/tree/main/optillm/plugins/deep_research
Sample reports: https://github.com/codelion/optillm/tree/main/optillm/plugins/deep_research/sample_reports
Thanks to the authors for this innovative approach! The application of diffusion concepts to text generation opens up exciting possibilities. Would love to collaborate on further improvements or applications of this technique.
The paper proposes Test-Time Diffusion Deep Researcher (TTD-DR). Unlike traditional AI research tools that generate content in one shot, TTD-DR uses an iterative refinement process:
• Starts with an initial "noisy" draft
• Performs gap analysis to identify missing information
• Actively searches the web to fill knowledge gaps
• Iteratively refines the report through multiple denoising steps
• Ensures all claims are backed by real web sources
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper