Sparsity-Guided Holistic Explanation for LLMs with Interpretable Inference-Time Intervention (AAAI'24)

Datasets and baselines are available in our prevous work: CBM-NLP

Abstract

Large Language Models (LLMs) have achieved unprecedented breakthroughs in various natural language processing domains. However, the enigmatic ``black-box'' nature of LLMs remains a significant challenge for interpretability, hampering transparent and accountable applications.

While past approaches, such as attention visualization, pivotal subnetwork extraction, and concept-based analyses, offer some insight, they often focus on either local or global explanations within a single dimension, occasionally falling short in providing comprehensive clarity.

In response, we propose a novel methodology anchored in sparsity-guided techniques, aiming to provide a holistic interpretation of LLMs.

Our framework, termed \textit{SparseCBM}, innovatively integrates sparsity to elucidate three intertwined layers of interpretation: input, subnetwork, and concept levels. In addition, the newly introduced dimension of interpretable inference-time intervention facilitates dynamic adjustments to the model during deployment.

Through rigorous empirical evaluations on real-world datasets, we demonstrate that SparseCBM delivers a profound understanding of LLM behaviors, setting it apart in both interpreting and ameliorating model inaccuracies. Codes are provided in supplements.

Install

We follow installation instructions from the CEBaB repository, which mainly depends on Huggingface.

Experiments

The code is tested on NVIDIA A100 80GB GPUs. An example for running the experiments is as follows:

bash run.sh

Note: It seems the random seed cannot control the randomness in parameter initialization in transformer, we suggest to run the code multiple times to get good scores.

Citation

@article{tan2023sparsity,
  title={Sparsity-Guided Holistic Explanation for LLMs with Interpretable Inference-Time Intervention},
  author={Tan, Zhen and Chen, Tianlong and Zhang, Zhenyu and Liu, Huan},
  journal={arXiv preprint arXiv:2312.15033},
  year={2023}
}