This is the official repository for "Mechanistic Interpretability of Emotion Inference in Large Language Models"

Requirements

Please see requirements.txt

Experiments

The main file to run experiments is main.py. For example, to run the probing experiment on Llama 3.2 1B, run:

python main.py --model_index=0 --in-domain --extract_hidden_states --emotion_probing

where the first argument just filters the samples that the model predicted correctly, the second argument extracts the activations and the last argument applies probing and save the results.

Visualization

To visualize the results of the experiments, use one of the plotters_x.ipynb notebooks provided as following: &rarr

plotters_1.ipynb : To plot results for a specific language model, e.g. Llama 3.2 1B
plotters_2.ipynb : To visualize results for all language models at the same time.
plotters_3.ipynb : To study effect of prompt template
plotters_4.ipynb : To study the isomorphic control task

📄 Citation

Our paper is accepted to ACL 2025 in Vienna Austria.
You can find it in the official program here:
🔗 https://2025.aclweb.org/program/find_papers/

If you found our paper helpful, please cite us as:

@inproceedings{tak2025mechanistic,
  title     = {Mechanistic Interpretability of Emotion Inference in Large Language Models},
  author    = {Tak, Ala N. and Banayeeanzade, Amin and Bolourani, Anahita and Kian, Mina and Jia, Robin and Gratch, Jonathan},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2025},
  month     = jul,
  year      = {2025},
  address   = {Vienna, Austria},
  publisher = {Association for Computational Linguistics},
  pages     = {13090--13120}
}