emnlp

Fints

This is the repo for Fints: Efficient Inference-Time Personalization for LLMs with Fine-Grained Instance-Tailored Steering.

<img width="659" height="260" alt="image" src="https://github.com/user-attachments/assets/3828ccd0-e96f-42b2-ab9e-5b95968b160c" />

Data Preparation

See under [data_process/]. The preparation process is for Headline Generation, Abstract Writing, and PersonalWAB

Run

./new_data_process.py

to select some users for experiments. Run

./ranking.py --task [your_task]

to sort all historical samples according to their relevance to the corresponding training or testing samples.

Negative Samples Generation

For PersonalWAB, run

python data_collect_pwab.py --dataset pwab \
--modelweight [root_of_models] \
--k 5 \
--llm llama-3.1 \
--form json \
--data_path ../pa_back/data

to generate positive samples.

For all datasets, run

./data_collect.sh

to generate negtive samples.

Vectors Generation and Evaluation

Run

./run.sh

to generate personalized steering vectors and evaluate the model on the test set.

After evaluation, for PersonalWAB, the evaluation scores will be saved; for other datasets, add the parameter --eval in run_generation.py and run again to obtain the model's evaluation scores.

By adding --plugin in run_generation.py, PA-steering can be used with lora models.

Citation

If you find this repo useful, please cite our paper:

@misc{du2025fintsefficientinferencetimepersonalization,
      title={Fints: Efficient Inference-Time Personalization for LLMs with Fine-Grained Instance-Tailored Steering}, 
      author={Kounianhua Du and Jianxing Liu and Kangning Zhang and Wenxiang Jiao and Yuan Lu and Jiarui Jin and Weiwen Liu and Yong Yu and Weinan Zhang},
      year={2025},
      eprint={2510.27206},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2510.27206}, 
}