Saved in:
Bibliographic Details
Main Authors: Li, Jiahang, Chen, Taoyu, Wang, Yuanli
Format: Preprint
Published: 2023
Subjects:
Online Access:https://arxiv.org/abs/2401.02976
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909063310213120
author Li, Jiahang
Chen, Taoyu
Wang, Yuanli
author_facet Li, Jiahang
Chen, Taoyu
Wang, Yuanli
contents This study introduces a novel approach for analyzing and modifying entity relationships in GPT models, diverging from ROME's entity-focused methods. We develop a relation tracing technique to understand the influence of language model computations on relationship judgments. Using the FewRel dataset, we identify key roles of MLP modules and attention mechanisms in processing relationship information. Our method, tested against ROME on a new dataset, shows improved balance in specificity and generalization, underscoring the potential of manipulating early-layer modules for enhanced model understanding and accuracy.
format Preprint
id arxiv_https___arxiv_org_abs_2401_02976
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Trace and Edit Relation Associations in GPT
Li, Jiahang
Chen, Taoyu
Wang, Yuanli
Computation and Language
Artificial Intelligence
This study introduces a novel approach for analyzing and modifying entity relationships in GPT models, diverging from ROME's entity-focused methods. We develop a relation tracing technique to understand the influence of language model computations on relationship judgments. Using the FewRel dataset, we identify key roles of MLP modules and attention mechanisms in processing relationship information. Our method, tested against ROME on a new dataset, shows improved balance in specificity and generalization, underscoring the potential of manipulating early-layer modules for enhanced model understanding and accuracy.
title Trace and Edit Relation Associations in GPT
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2401.02976