ReZero: Enhancing LLM search ability by trying one-more-time

ReZero trains a small language model to develop effective search behaviors instead of memorizing static data. It interacts with multiple synthetic search engines, each with unique retrieval mechanisms, to refine queries and persist in searching until it finds exact answers. The project focuses on reinforcement learning, preventing overfitting, and optimizing for efficiency in real-world search applications.

Quick Demo | Setup | Data and Training | Models | Experiments | References | Acknowledgements

Setup 🛠️

# Clone the repository
git clone https://github.com/menloresearch/ReZero
cd ReZero

# Create virtual environment
python -m venv .venv

# Activate the environment
source .venv/bin/activate

# Install dependencies
pip install --upgrade pip
pip install -e .

# Set up environment variables (required for websearch demo)
cp .env.example .env
# Edit .env and add your Tavily API key if you want to use the websearch demo

Quick Demo 🚀

Run the interactive web interface to see ReZero in action:

python app.py

This will launch a Gradio interface where you can interact with the model and test different search behaviors.

Data and Training 🧠

All necessary training data is included in the data/ folder. To train:

python train_grpo.py

If you want to regenerate the data, please run:

python scripts/generate_data.py

Models 🤖

You can find our models on Hugging Face 🤗! We're committed to open-source and easy access for the research community.

Model	Backbone	Size	Link	GGUF
ReZero-v0.1	Llama-3.2-3B	3B	🤗 Menlo/ReZero-v0.1-llama-3.2-3b-it-grpo-250404	🤗 GGUF

Experiments 🧪

Run ID	Model Config	Dataset	Steps	Hardware	TensorBoard	Description
exp-01	Llama-3.2-3b-instruct	Apollo Mission Report	300	~2 hours on 1xH200	📊	Added reward_search_strategy and reward_search_quality. Reward weights: [4.0, 2.0, 1.0, 1.0, 1.0, 1.0]. Loss crashed after step 400. Best accuracy: 31.25% at step 400. Max agent turns: 10.
exp-02	Llama-3.2-3b-instruct	Apollo Mission Report	1000	~7 hours on 1xH200	📊	Improved reward_retry logic to only reward search when answers found. Increased max agent turns to 20. Reward weights: [4.0, 2.0, 1.0, 1.0, 1.0, 1.0]. Best accuracy: 46.88% at step 250. Higher early reward_correctness (~0.6 vs 0.4-0.5). Loss stable but reward crashed after step 350.
exp-03	Llama-3.2-3b-instruct	Apollo Mission Report	1000	~7 hours on 1xH200	📊	Same as exp-02 but without the retry reward function.

References 📖

Acknowledgements 🤝

• This project is kickstarted from the source code of AutoDidact