Fetch

Code for the paper Don't Get Lost in the Trees: Streamlining LLM Reasoning by Overcoming Tree Search Exploration Pitfalls

---

🚀 Setup

Follow the steps below to run our scripts:

📌 Step 1. Setup service of policy, verifier, and embedding model

📚 Policy

We employ vllm for the policy. To start the policy service, run the following command:

python3 -m vllm.entrypoints.openai.api_server --model /path/to/policy/model --port 8000 --dtype float16 --tensor-parallel-size 2 --swap-space 8 --max-model-len 4096

🔍 Verifier

1. Update your model path in verifier/server.py.
2. Run the script: bash run.sh ./ 0 inside the verifier directory.

📦 Embedding Model

If you're using state merging, follow these steps:

1. Update the path in cluster/server_cluster.py.
2. Run the script: bash run_app.sh ./ 0 inside the cluster directory.

---

📌 Step 2. Run tree search algorithms

We provide three tree search algorithms: BFS (Best-First Search), Beam Search, and MCTS (Monte Carlo Tree Search).

1. Specify the input, output file paths, and other parameters in scripts such as beamsearch.py.

2. Simply execute the corresponding Python script. For instance, to run Beam Search: python3 beamsearch.py

---

🎯 Tips

• Using Your Own Models: If you prefer to use your own policy or verifier, please pay attention to the prompts fed to the models and the separator (e.g., "\n") to split steps.
• Embedding Model Preparation: Preparing an embedding model can improve performance. We provide scripts to train these models. You can also use pretrained SimCSE or other embedding models available on Hugging Face Hub.
• Pre-trained Models: If you don't want to train models, our trained models are available on HuggingFace Hub:
  • Policy: xmu-nlp/Llama-3-8b-gsm8k
  • Verifier: xmu-nlp/Llama-3-8b-gsm8k-value-A and xmu-nlp/Llama-3-8b-gsm8k-value-B
  • Emb: xmu-nlp/simcse-large-gsm8k

---

📝 Citation

If you find our work useful, please cite our paper:

@misc{wang2025dontlosttreesstreamlining,
      title={Don't GetLost in the Trees: Streamlining LLM Reasoning by Overcoming Tree Search Exploration Pitfalls}, 
      author={Ante Wang and Linfeng Song and Ye Tian and Dian Yu and Haitao Mi and Xiangyu Duan and Zhaopeng Tu and Jinsong Su and Dong Yu},
      year={2025},
      eprint={2502.11183},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.11183}, 
}