Saved in:
Bibliographic Details
Main Authors: Abbasi, Sina, Modarres, Mohammad Reza, Pilehvar, Mohammad Taher
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2411.16252
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913994378313728
author Abbasi, Sina
Modarres, Mohammad Reza
Pilehvar, Mohammad Taher
author_facet Abbasi, Sina
Modarres, Mohammad Reza
Pilehvar, Mohammad Taher
contents With new large language models (LLMs) emerging frequently, it is important to consider the potential value of model-agnostic approaches that can provide interpretability across a variety of architectures. While recent advances in LLM interpretability show promise, many rely on complex, model-specific methods with high computational costs. To address these limitations, we propose NormXLogit, a novel technique for assessing the significance of individual input tokens. This method operates based on the input and output representations associated with each token. First, we demonstrate that during the pre-training of LLMs, the norms of word embeddings effectively capture token importance. Second, we reveal a significant relationship between a token's importance and the extent to which its representation can resemble the model's final prediction. Extensive analyses reveal that our approach outperforms existing gradient-based methods in terms of faithfulness and offers competitive performance in layer-wise explanations compared to leading architecture-specific techniques.
format Preprint
id arxiv_https___arxiv_org_abs_2411_16252
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle NormXLogit: The Head-on-Top Never Lies
Abbasi, Sina
Modarres, Mohammad Reza
Pilehvar, Mohammad Taher
Computation and Language
With new large language models (LLMs) emerging frequently, it is important to consider the potential value of model-agnostic approaches that can provide interpretability across a variety of architectures. While recent advances in LLM interpretability show promise, many rely on complex, model-specific methods with high computational costs. To address these limitations, we propose NormXLogit, a novel technique for assessing the significance of individual input tokens. This method operates based on the input and output representations associated with each token. First, we demonstrate that during the pre-training of LLMs, the norms of word embeddings effectively capture token importance. Second, we reveal a significant relationship between a token's importance and the extent to which its representation can resemble the model's final prediction. Extensive analyses reveal that our approach outperforms existing gradient-based methods in terms of faithfulness and offers competitive performance in layer-wise explanations compared to leading architecture-specific techniques.
title NormXLogit: The Head-on-Top Never Lies
topic Computation and Language
url https://arxiv.org/abs/2411.16252