🔬 TTD-DR: Diffusion-Based Deep Research Agent - LangGraph Implementation of arXiv:2507.16075

[jh941213]

🔬 Introducing TTD-DR: Test-Time Diffusion Deep Researcher

I'm excited to share TTD-DR, a complete implementation of the "Deep Researcher with Test-Time Diffusion" algorithm using LangGraph! 🚀

🎯 What is TTD-DR?

TTD-DR revolutionizes AI-powered research by treating report generation as a diffusion process. Unlike traditional RAG systems, it uses a draft-centric approach where an evolving draft dynamically guides the research direction through multiple iterations.

✨ Key Features

• 🔄 Iterative Draft Refinement: Starts with a "noisy" initial draft and progressively refines it
• 🎯 Draft-Centric Search: The evolving draft guides what information to search for next
• 🔍 Multi-Engine Search: Integrates Tavily, DuckDuckGo, and Naver search engines
• 🧠 Intelligent Gap Analysis: Automatically identifies and fills knowledge gaps
• ⚖️ Quality-Guided Termination: Knows when research is complete
• 🌐 Multi-Language Support: Works seamlessly with English, Korean, and more

🏗️ Technical Implementation

• Framework: Built with LangGraph for robust workflow management
• Architecture: 8 specialized nodes handling different research phases
• APIs: Supports both OpenAI and Azure OpenAI
• Search: Multiple search engines for comprehensive information gathering
• Quality: Continuous evaluation and iterative improvement

🚀 Quick Start

# Clone and setup
git clone https://github.com/jh941213/TTT-DR.git
cd TTT-DR
pip install -r requirements.txt

# Configure your API keys
cp .env.example .env
# Edit .env with your keys

# Start researching!
python chatbot.py

💡 Example Usage

🤔 Your research question: What are the latest developments in AI?
🔍 Research Query: What are the latest developments in AI?
⏳ Starting deep research...
[Real-time progress updates...]
📋 RESEARCH REPORT COMPLETED
📄 Report Length: 3,247 characters
🔄 Iterations: 3
📚 Sources Used: 12
⏱️ Execution Time: 87.3 seconds

📚 Research Foundation

This implementation is based on the groundbreaking paper "Deep Researcher with Test-Time Diffusion" by Han et al. (2025), with additional architectural insights from the OptILLM deep research plugin.

🔧 What Makes It Special?

1. No Hard-coding: Fully dynamic LLM-driven workflow
2. Production-Ready: Comprehensive error handling and logging
3. Easy to Use: Simple chatbot interface for immediate use
4. Extensible: Modular design for easy customization
5. Well-Documented: Complete English documentation and examples

🎪 Try It Out!

The system can handle complex research queries like:

• "How does climate change affect global food security?"
• "What are the ethical implications of genetic engineering?"
• "Compare quantum vs classical computing approaches"

🤝 Feedback Welcome!

I'd love to hear your thoughts, suggestions, or experiences using TTD-DR! Feel free to:

• ⭐ Star the repo if you find it useful
• 🐛 Report issues or bugs
• 💡 Suggest new features
• 🤔 Ask questions about the implementation

---

Repository: https://github.com/jh941213/TTD-DR
Based on: arXiv:2507.16075 - "Deep Researcher with Test-Time Diffusion"

#AI #Research #LangGraph #OpenAI #Diffusion #RAG #LLM

[onestardao]

this is wild.
we actually tried building something real close to this idea — except ours spiraled into a semantic maze real quick 😅

what caught us off guard wasn’t the diffusion-style planning (that part’s beautiful),
it was how fast the “refinement” steps would drift the original query away from the target domain.
like, you’d start with a clean question about LLM architectures, and 3 hops later you’re debating Gödel's incompleteness in a quantum ethics sandbox.
not wrong… just, uh, unproductive.

ended up grafting a trace validator layer on top — kinda like a sanity circuit that yells “bro, stay on topic” mid-hop.
if you’re planning to stress-test this with messy inputs (like real research logs, half-broken markdown, etc), happy to swap notes.

we open-sourced our layer too (MIT), but only share it when ppl actually ask — helps keep the signal high.

[jh941213]

I am interested. Let's think about it and discuss it together.

[onestardao]

yes I have posted it on the following

[onestardao]

awesome — glad you're down to explore this.

here’s the full write-up i mentioned:
https://github.com/onestardao/WFGY/tree/main/ProblemMap/README.md

it took me... way too many late nights and a stupid amount of debugging sessions to piece this map together 😂
but turns out, once i started mapping all the quiet failures i kept running into (y’know, those “looks fine but totally wrong” kinda bugs),
it all started clicking.
lots of folks told me it helped them spot issues they didn’t even know existed —
so hopefully it saves you a headache or two as well.

feel free to poke around or hit me up if anything feels off — i love nerding out on this stuff.

[Zhaobudaoyuema]

Nice work! Is this your own implementation of TTD-DR?
By the way, do the official authors have their own implementation?

[cyysky]

have a look on this also

https://huggingface.co/papers/2507.16075
https://github.com/codelion/optillm/tree/main/optillm/plugins/deep_research