Test-Time Learning for Large Language Models

Jinwu Hu, Zitian Zhang, Guohao Chen, Xutao Wen, Chao Shuai, Wei Luo, Bin Xiao, Yuanqing Li, Mingkui Tan
_{South China University of Technology, Pazhou Laboratory, Zhejiang University, South China Agricultural University, Chongqing University of Posts and Telecommunications}

🔥News

• 2025-07-31: Update AdaptEval benchmark and models.
• 2025-05-27: We have released our paper on Arxiv.
• 2025-05-01: TLM is accepted by ICML2025.

🚀Quick Start

## clone our repo
git clone https://github.com/Fhujinwu/TLM.git
cd TLM
## install TLM environment
conda create --name tlm --yes python=3.10
conda activate tlm
pip install -e ".[torch,metrics]" --no-build-isolation

🗂 Benchmarks and models

• Benchmarks：https://huggingface.co/datasets/Jinwu01/AdaptEval
• Models: https://huggingface.co/Jinwu01/TLM

🔨 Training

All datasets and their contents from AdaptEval are defined in the dataset_info.json file included in this repository. You only need to specify the desired dataset in your configuration file to use it.

For example, to adapt to the geography dataset:

• For offline test-time learning, you can start training with the following command:

CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/train_lora/offline_ttl.yaml

• For online test-time learning, use:

CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/train_lora/online_ttl.yaml

The offline_ttl.yaml and online_ttl.yaml files provide example configurations for fine-tuning with test-time learning. These configurations specify parameters about model, fine-tuning method, dataset, TTL method and so on. Please customize these files according to your own requirements.

⚖️ Evaluation

After running the above training commands, you will obtain the model inference results in the specified output_dir. You can then evaluate these results.

First, install the required dependencies:

pip install rouge_score rouge-chinese bert_score git+https://github.com/google-research/bleurt.git

All evaluation-related scripts are located in the scripts/eval folder:

• For datasets in DomainBench and InstructionBench, copy the path to your model inference results into eval_simility.py and run the script.
• For datasets in ReasoningBench, copy the path to your model inference results into eval_accuracy.py and run the script.

💬 Citation

Thanks for the open-source code of LLaMA-Factory

If you find our work interesting and meaningful, welcome to give a 🌟 to our repo and cite our paper.

@inproceedings{hutest,
  title={Test-Time Learning for Large Language Models},
  author={Hu, Jinwu and Zhang, Zitian and Chen, Guohao and Wen, Xutao and Shuai, Chao and Luo, Wei and Xiao, Bin and Li, Yuanqing and Tan, Mingkui},
  booktitle={Forty-second International Conference on Machine Learning}
}

Star History