ReasonGraph

What is ReasonGraph ?

ReasonGraph is an open source LLM (large language model) inference visualization and analysis tool. It transforms LLM's complex reasoning process into clear visual charts, helping users understand AI's decision logic and optimize inference paths.

• Supports 50+ advanced LLMs (such as OpenAI, Google, Anthropic, DeepSeek, etc.)

• Compatible with multiple inference methods (sequential reasoning, tree inference, etc.)

• Interactive visualization (real-time update of inference paths, supporting scaling, parameter adjustment, SVG export)

• Modular design (supports API extensions to new models and inference methods)

• Multilingual support (currently supported in Chinese and English)

GitHub open source address : https://github.com/ZongqianLi/ReasonGraph
Experience Demo online : https://huggingface.co/spaces/ZongqianLi/ReasonGraph
Technical paper (arXiv) : https://arxiv.org/pdf/2503.03979

Main functions

✅ Inference path visualization: convert text inference process into intuitive flowcharts 

✅ Various inference methods: support sequential reasoning, tree inference, self-improvement, beam search, etc. 

✅ Real-time interaction: support dynamic updates, scaling adjustments, parameter optimization 

✅ Long inference support: can handle long text inference paths generated by LLM

✅Open API interface : expandable new models and inference methods

Technical Architecture

• Front-end visualization : implement dynamic graphics rendering based on Mermaid.js

• Backend framework : Use Flask , including the following core modules:

• Configuration Manager : Manage status updates and configurations

• API Factory : Provides a unified API, compatible with multiple LLMs

• Reasoning Methods Module : encapsulate different inference methods to realize standardized analysis and visualization

• Inference path analysis : parsing LLM output based on XML, converting to tree or directed graph

• Real-time update : supports user adjustment parameters, backend calls LLM instantly and updates visual charts

• Supports multi-platform deployment : available on-premises, cloud or Hugging Face

Applicable scenarios

Academic research: Visualizing the reasoning process, analyzing the performance of LLM in complex tasks Education and teaching: Helping students understand AI logic, improve learning efficiency Model optimization: analyzing LLM's reasoning errors, and optimizing the reasoning path
AI application development: Support logical optimization of LLM applications and improve user experience research on reasoning methods: Exploring new AI reasoning strategies and promoting technological innovation

Quick experience

1️⃣Access online Demo : https://huggingface.co/spaces/ZongqianLi/ReasonGraph
2️⃣ Install the local environment :

 pip install -r requirements.txt

3️⃣Run Flask server :

 python app.py

4️⃣Access local services : http://127.0.0.1:5000

Latest updates

2025.03.24: Support long text reasoning visualization!
2025.03.18: Add more LLM compatible models, supporting Chinese and English
2025.03.17: Hugging Face launches online experience demo
2025.03.07: Technical papers are published on arXiv
2025.02.22: GitHub project creation