Official repository for Meta-Reasoning Improves Tool Use in Large Language Models

Instructions

Before running the repository, the Llama3 8B Instruct checkpoint must be obtained. See instructions here.

Installation:

git clone https://github.com/lisaalaz/tecton.git
cd tecton
python -m venv 'tecton_env'
. tecton_env/bin/activate
pip install -r requirements.txt

To fine-tune on GSM8K-XL:

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.run --nproc_per_node 1 --master_port 1200 train.py \
--ckpt_dir llama3/8B \
--tokenizer_path llama3/8B/tokenizer.model \
--input_file data/gsm8k-xl/train.json \
--lr 1e-3 \
--num_epochs 10 \
--dataset gsm8k-xl

To run inference on GSM8K-XL with TECTON-score:

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.run --nproc_per_node 1 --master_port 3000 inference.py \
--ckpt_dir llama3/8B-instruct \
--tokenizer_path llama3/8B-instruct/tokenizer.model \
--mode score \
--dataset gsm8k-xl \
--func_load_path checkpoints/best_checkpoint.pth \
--logits_bias 3.0

To run inference on GSM8K-XL with TECTON-generate:

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.run --nproc_per_node 1 --master_port 3000 inference.py \
--ckpt_dir llama3/8B-instruct \
--tokenizer_path llama3/8B-instruct/tokenizer.model \
--mode generate \
--dataset gsm8k-xl \
--func_load_path checkpoints/best_checkpoint.pth \
--logits_bias 3.0

Citation

@inproceedings{alazraki-rei-2025-meta,
    title = "Meta-Reasoning Improves Tool Use in Large Language Models",
    author = "Alazraki, Lisa and Rei, Marek",
    editor = "Chiruzzo, Luis and Ritter, Alan and Wang, Lu",
    booktitle = "Findings of the Association for Computational Linguistics: NAACL 2025",
    month = apr,
    year = "2025",
    address = "Albuquerque, New Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.findings-naacl.440/",
    doi = "10.18653/v1/2025.findings-naacl.440",
    pages = "7900--7912",
    ISBN = "979-8-89176-195-7",
}