InnoTree: Integrating MCTS with Language Models for Scientific Idea Generation

💡 InnoTree is an advanced agent system specifically designed to integrate Monte Carlo Tree Search (MCTS) techniques, enabling the efficient generation and exploration of innovative scientific ideas.

🤗 Introduction

We introduce InnoTree, an agent system designed to facilitate the generation and exploration of innovative scientific ideas. InnoTree integrates Monte Carlo Tree Search (MCTS) techniques to efficiently navigate complex idea spaces, systematically identifying and evaluating promising research directions. By leveraging the probabilistic nature of MCTS, the system strikes a balance between exploration and exploitation, ensuring a comprehensive yet focused search for high-impact scientific concepts. This framework offers a structured approach to idea development, with considerable potential to accelerate interdisciplinary research and foster novel discoveries. Experimental results demonstrate that InnoTree generates more innovative scientific ideas compared to traditional methods.

🌟 Quick Start

Cloning the repo

git clone https://github.com/Goer17/InnoTree.git

Downloading the dependencies

conda create -n inno_tree python=3.12 -y && \
conda activate inno_tree

pip install -r requirements.txt

Running

Create a .env file and set your API key :

CUSTOM_API_KEY=
CUSTOM_BASE_URL=

OPENAI_API_KEY=
OPENAI_BASE_URL=https://api.openai.com/v1

GROK_API_KEY=
GROK_BASE_URL=https://api.grok.com/v1

DEEPSEEK_API_KEY=
DEEPSEEK_BASE_URL=https://api.deepseek.com

# ...

python main.py --topic "Multi Agent System" \ # choose the target topic
    --model "gpt-4o-mini" \
    --n_rollouts 10 \
    --n_trials 20 \
    --n_exp 3 \
    --n_results 5 \
    --arena \
    --sampling best

Client-Server Model

Alternatively, you can run the application in a client-server model :

python app.py

Then, use curl to send a POST request to the server :

curl -X POST http://127.0.0.1:5000/start \
     -H "Content-Type: application/json" \
     -d '{
           "topic": "Target topic",
           "api_key": "Your API key",
           "base_url": "Your base url",
           "model": "Your model",
           "sampling_method": "best", # Decoding method: best, epsilon, v-epsilon
           "exploration_weight": 1.5,
           "n_trials": 15,
           "n_rollouts": 12,
           "n_expand": 3
         }'

You can retrieve your task ID from the response and then use it to stream the results :

curl -X GET http://127.0.0.1:5000/stream\?task_id\="Your task id"

👀 Visualization

We’ve also developed a lightweight GUI using Vue.js to visualize the search path of InnoTree. This interface provides insight into how the Monte Carlo Tree Search (MCTS) algorithm operates, showing how it explores and updates the value of each node during the process.

To set up the front-end, navigate to the inno-tree-web directory and install the required dependencies :

cd inno-tree-web && npm install

Start the front-end development server:

npm run dev

To run the back-end server, navigate back to the root directory and start the application :

cd .. && python app.py

Once everything is running, you can access the application at http://127.0.0.1:5173