ExplanationIntervention

Can Language Models Teach? Teacher Explanations Improve Student Performance via Personalization

Swarnadeep Saha, Peter Hase, and Mohit Bansal

Overview of single-round of interaction between a teacher LLM and a student LLM, covering first three research questions

Overview of multi-round student-teacher interaction, detailing the fourth research question

Installation

This repository is tested on Python 3.10.11.
You should install this repository on a virtual environment. All dependencies can be installed as follows:

pip install -r requirements.txt

Datasets

The datasets are already included in datasets folder.

Research Question 1

In order to get results for RQ1, first update the cache directory where pre-trained models will be saved. Then execute the following commands.

cd src
python main_single_turn.py --intervention_strategy random --results_file ../results/rq1.txt

This, by default will run experiments on the StrategyQA dataset with the Flan-T5 models. Update the model and dataset paths accordingly.

Research Question 2

Execute the following commands.

cd src
python main_single_turn.py --intervention_strategy mm_both --results_file ../results/rq2.txt

Research Question 3

Execute the following commands.

cd src
python main_single_turn.py --intervention_strategy mm_both --teacher_expl_type useful_teacher --results_file ../results/rq3.txt

Research Question 4

Execute the following commands.

cd src
python main_multi_turn.py --results_file ../results/rq4.txt

Research Question 5

Execute the following commands.

cd src
python main_single_turn.py --deceive True --results_file ../results/rq5.txt

The harmful/deceiving explanations we constructed (two per sample) are placed in datasets/strategyqa_dataset/harmful_explanations.json.

Citation

@inproceedings{saha2023can,
  title={Can Language Models Teach Weaker Agents? Teacher Explanations Improve Students via Personalization},
  author={Saha, Swarnadeep and Hase, Peter and Bansal, Mohit},
  booktitle={NeurIPS},
  url={https://arxiv.org/abs/2306.09299},
  year={2023}
}