Detecting Edit Failures In Large Language Models: An Improved Specificity Benchmark

Jason Hoelscher-Obermaier; Julia H. Persson; Esben Kran; Ioannis Konstas; Fazl Barez

doi:10.18653/v1/2023.findings-acl.733

Detecting Edit Failures In Large Language Models: An Improved Specificity Benchmark

Jason Hoelscher-Obermaier
, Julia H. Persson
, Esben Kran
, Ioannis Konstas
, Fazl Barez^*

^*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

27 Citations (Scopus)

1 Downloads (Pure)

Abstract

Recent model editing techniques promise to mitigate the problem of memorizing false or outdated associations during large language model (LLM) training. However, we show that these techniques can introduce large unwanted side effects which are not detected by existing specificity benchmarks. We extend the existing COUNTERFACT benchmark to include a dynamic component and dub our benchmark COUNTERFACT+. Additionally, we extend the metrics used for measuring specificity by a principled KL divergence-based metric. We use this improved benchmark to evaluate recent model editing techniques and find that they suffer from low specificity. Our findings highlight the need for improved specificity benchmarks that identify and prevent unwanted side effects.

Original language	English
Title of host publication	Findings of the Association for Computational Linguistics, ACL 2023
Publisher	Association for Computational Linguistics
Pages	11548-11559
Number of pages	12
ISBN (Electronic)	9781959429623
DOIs	https://doi.org/10.18653/v1/2023.findings-acl.733
Publication status	Published - 9 Jul 2023
Event	61st Annual Meeting of the Association for Computational Linguistics 2023 - Toronto, Canada Duration: 9 Jul 2023 → 14 Jul 2023

Conference

Conference	61st Annual Meeting of the Association for Computational Linguistics 2023
Abbreviated title	ACL 2023
Country/Territory	Canada
City	Toronto
Period	9/07/23 → 14/07/23

ASJC Scopus subject areas

Computer Science Applications
Linguistics and Language
Language and Linguistics

Access to Document

10.18653/v1/2023.findings-acl.733Licence: CC BY

Download pdf
©2023 Association for Computational Linguistics ACL materials are Copyright © 1963–2025 ACL; other materials are copyrighted by their respective copyright holders. Materials prior to 2016 here are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 International License. Permission is granted to make copies for the purposes of teaching and research. Materials published in or after 2016 are licensed on a Creative Commons Attribution 4.0 International License.
Final published version, 603 KBLicence: CC BY

Cite this

@inproceedings{d7d3f761d7ab4d9384eae86909b57b1b,

title = "Detecting Edit Failures In Large Language Models: An Improved Specificity Benchmark",

abstract = "Recent model editing techniques promise to mitigate the problem of memorizing false or outdated associations during large language model (LLM) training. However, we show that these techniques can introduce large unwanted side effects which are not detected by existing specificity benchmarks. We extend the existing COUNTERFACT benchmark to include a dynamic component and dub our benchmark COUNTERFACT+. Additionally, we extend the metrics used for measuring specificity by a principled KL divergence-based metric. We use this improved benchmark to evaluate recent model editing techniques and find that they suffer from low specificity. Our findings highlight the need for improved specificity benchmarks that identify and prevent unwanted side effects.",

author = "Jason Hoelscher-Obermaier and Persson, \{Julia H.\} and Esben Kran and Ioannis Konstas and Fazl Barez",

note = "Publisher Copyright: {\textcopyright} 2023 Association for Computational Linguistics.; 61st Annual Meeting of the Association for Computational Linguistics 2023, ACL 2023 ; Conference date: 09-07-2023 Through 14-07-2023",

year = "2023",

month = jul,

day = "9",

doi = "10.18653/v1/2023.findings-acl.733",

language = "English",

pages = "11548--11559",

booktitle = "Findings of the Association for Computational Linguistics, ACL 2023",

publisher = "Association for Computational Linguistics",

}

Hoelscher-Obermaier, J, Persson, JH, Kran, E, Konstas, I & Barez, F 2023, Detecting Edit Failures In Large Language Models: An Improved Specificity Benchmark. in Findings of the Association for Computational Linguistics, ACL 2023. Association for Computational Linguistics, pp. 11548-11559, 61st Annual Meeting of the Association for Computational Linguistics 2023, Toronto, Ontario, Canada, 9/07/23. https://doi.org/10.18653/v1/2023.findings-acl.733

Detecting Edit Failures In Large Language Models: An Improved Specificity Benchmark. / Hoelscher-Obermaier, Jason; Persson, Julia H.; Kran, Esben et al.
Findings of the Association for Computational Linguistics, ACL 2023. Association for Computational Linguistics, 2023. p. 11548-11559.

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

TY - GEN

T1 - Detecting Edit Failures In Large Language Models: An Improved Specificity Benchmark

AU - Hoelscher-Obermaier, Jason

AU - Persson, Julia H.

AU - Kran, Esben

AU - Konstas, Ioannis

AU - Barez, Fazl

PY - 2023/7/9

Y1 - 2023/7/9

N2 - Recent model editing techniques promise to mitigate the problem of memorizing false or outdated associations during large language model (LLM) training. However, we show that these techniques can introduce large unwanted side effects which are not detected by existing specificity benchmarks. We extend the existing COUNTERFACT benchmark to include a dynamic component and dub our benchmark COUNTERFACT+. Additionally, we extend the metrics used for measuring specificity by a principled KL divergence-based metric. We use this improved benchmark to evaluate recent model editing techniques and find that they suffer from low specificity. Our findings highlight the need for improved specificity benchmarks that identify and prevent unwanted side effects.

AB - Recent model editing techniques promise to mitigate the problem of memorizing false or outdated associations during large language model (LLM) training. However, we show that these techniques can introduce large unwanted side effects which are not detected by existing specificity benchmarks. We extend the existing COUNTERFACT benchmark to include a dynamic component and dub our benchmark COUNTERFACT+. Additionally, we extend the metrics used for measuring specificity by a principled KL divergence-based metric. We use this improved benchmark to evaluate recent model editing techniques and find that they suffer from low specificity. Our findings highlight the need for improved specificity benchmarks that identify and prevent unwanted side effects.

UR - https://www.scopus.com/pages/publications/85166307252

U2 - 10.18653/v1/2023.findings-acl.733

DO - 10.18653/v1/2023.findings-acl.733

M3 - Conference contribution

AN - SCOPUS:85166307252

SP - 11548

EP - 11559

BT - Findings of the Association for Computational Linguistics, ACL 2023

PB - Association for Computational Linguistics

T2 - 61st Annual Meeting of the Association for Computational Linguistics 2023

Y2 - 9 July 2023 through 14 July 2023

ER -

Detecting Edit Failures In Large Language Models: An Improved Specificity Benchmark

Abstract

Conference

ASJC Scopus subject areas

Access to Document

Other files and links

Fingerprint

Cite this