Reasoning Robustness of LLMs to Adversarial Typographical Errors

Large Language Models (LLMs) have demonstrated impressive capabilities in reasoning using Chain-of-Thought (CoT) prompting. However, CoT can be biased by users' instruction. In this work, we study the reasoning robustness of LLMs to typographical errors, which can naturally occur in users'...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Gan, Esther, Zhao, Yiran, Cheng, Liying, Mao, Yancan, Goyal, Anirudh, Kawaguchi, Kenji, Kan, Min-Yen, Shieh, Michael
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page
container_issue
container_start_page
container_title
container_volume
creator Gan, Esther
Zhao, Yiran
Cheng, Liying
Mao, Yancan
Goyal, Anirudh
Kawaguchi, Kenji
Kan, Min-Yen
Shieh, Michael
description Large Language Models (LLMs) have demonstrated impressive capabilities in reasoning using Chain-of-Thought (CoT) prompting. However, CoT can be biased by users' instruction. In this work, we study the reasoning robustness of LLMs to typographical errors, which can naturally occur in users' queries. We design an Adversarial Typo Attack ($\texttt{ATA}$) algorithm that iteratively samples typos for words that are important to the query and selects the edit that is most likely to succeed in attacking. It shows that LLMs are sensitive to minimal adversarial typographical changes. Notably, with 1 character edit, Mistral-7B-Instruct's accuracy drops from 43.7% to 38.6% on GSM8K, while with 8 character edits the performance further drops to 19.2%. To extend our evaluation to larger and closed-source LLMs, we develop the $\texttt{R$^2$ATA}$ benchmark, which assesses models' $\underline{R}$easoning $\underline{R}$obustness to $\underline{\texttt{ATA}}$. It includes adversarial typographical questions derived from three widely used reasoning datasets-GSM8K, BBH, and MMLU-by applying $\texttt{ATA}$ to open-source LLMs. $\texttt{R$^2$ATA}$ demonstrates remarkable transferability and causes notable performance drops across multiple super large and closed-source LLMs.
doi_str_mv 10.48550/arxiv.2411.05345
format Article
fullrecord <record><control><sourceid>arxiv_GOX</sourceid><recordid>TN_cdi_arxiv_primary_2411_05345</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2411_05345</sourcerecordid><originalsourceid>FETCH-arxiv_primary_2411_053453</originalsourceid><addsrcrecordid>eNpjYJA0NNAzsTA1NdBPLKrILNMzMjE01DMwNTYx5WRwCEpNLM7Py8xLVwjKTyotLslLLS5WyE9T8PHxLVYoyVdwTClLLSpOLMpMzFEIqSzITy9KLMjITAbyXIuK8ouKeRhY0xJzilN5oTQ3g7yba4izhy7YrviCoszcxKLKeJCd8WA7jQmrAAD_rjbP</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>Reasoning Robustness of LLMs to Adversarial Typographical Errors</title><source>arXiv.org</source><creator>Gan, Esther ; Zhao, Yiran ; Cheng, Liying ; Mao, Yancan ; Goyal, Anirudh ; Kawaguchi, Kenji ; Kan, Min-Yen ; Shieh, Michael</creator><creatorcontrib>Gan, Esther ; Zhao, Yiran ; Cheng, Liying ; Mao, Yancan ; Goyal, Anirudh ; Kawaguchi, Kenji ; Kan, Min-Yen ; Shieh, Michael</creatorcontrib><description>Large Language Models (LLMs) have demonstrated impressive capabilities in reasoning using Chain-of-Thought (CoT) prompting. However, CoT can be biased by users' instruction. In this work, we study the reasoning robustness of LLMs to typographical errors, which can naturally occur in users' queries. We design an Adversarial Typo Attack ($\texttt{ATA}$) algorithm that iteratively samples typos for words that are important to the query and selects the edit that is most likely to succeed in attacking. It shows that LLMs are sensitive to minimal adversarial typographical changes. Notably, with 1 character edit, Mistral-7B-Instruct's accuracy drops from 43.7% to 38.6% on GSM8K, while with 8 character edits the performance further drops to 19.2%. To extend our evaluation to larger and closed-source LLMs, we develop the $\texttt{R$^2$ATA}$ benchmark, which assesses models' $\underline{R}$easoning $\underline{R}$obustness to $\underline{\texttt{ATA}}$. It includes adversarial typographical questions derived from three widely used reasoning datasets-GSM8K, BBH, and MMLU-by applying $\texttt{ATA}$ to open-source LLMs. $\texttt{R$^2$ATA}$ demonstrates remarkable transferability and causes notable performance drops across multiple super large and closed-source LLMs.</description><identifier>DOI: 10.48550/arxiv.2411.05345</identifier><language>eng</language><subject>Computer Science - Artificial Intelligence ; Computer Science - Computation and Language</subject><creationdate>2024-11</creationdate><rights>http://creativecommons.org/licenses/by/4.0</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>228,230,781,886</link.rule.ids><linktorsrc>$$Uhttps://arxiv.org/abs/2411.05345$$EView_record_in_Cornell_University$$FView_record_in_$$GCornell_University$$Hfree_for_read</linktorsrc><backlink>$$Uhttps://doi.org/10.48550/arXiv.2411.05345$$DView paper in arXiv$$Hfree_for_read</backlink></links><search><creatorcontrib>Gan, Esther</creatorcontrib><creatorcontrib>Zhao, Yiran</creatorcontrib><creatorcontrib>Cheng, Liying</creatorcontrib><creatorcontrib>Mao, Yancan</creatorcontrib><creatorcontrib>Goyal, Anirudh</creatorcontrib><creatorcontrib>Kawaguchi, Kenji</creatorcontrib><creatorcontrib>Kan, Min-Yen</creatorcontrib><creatorcontrib>Shieh, Michael</creatorcontrib><title>Reasoning Robustness of LLMs to Adversarial Typographical Errors</title><description>Large Language Models (LLMs) have demonstrated impressive capabilities in reasoning using Chain-of-Thought (CoT) prompting. However, CoT can be biased by users' instruction. In this work, we study the reasoning robustness of LLMs to typographical errors, which can naturally occur in users' queries. We design an Adversarial Typo Attack ($\texttt{ATA}$) algorithm that iteratively samples typos for words that are important to the query and selects the edit that is most likely to succeed in attacking. It shows that LLMs are sensitive to minimal adversarial typographical changes. Notably, with 1 character edit, Mistral-7B-Instruct's accuracy drops from 43.7% to 38.6% on GSM8K, while with 8 character edits the performance further drops to 19.2%. To extend our evaluation to larger and closed-source LLMs, we develop the $\texttt{R$^2$ATA}$ benchmark, which assesses models' $\underline{R}$easoning $\underline{R}$obustness to $\underline{\texttt{ATA}}$. It includes adversarial typographical questions derived from three widely used reasoning datasets-GSM8K, BBH, and MMLU-by applying $\texttt{ATA}$ to open-source LLMs. $\texttt{R$^2$ATA}$ demonstrates remarkable transferability and causes notable performance drops across multiple super large and closed-source LLMs.</description><subject>Computer Science - Artificial Intelligence</subject><subject>Computer Science - Computation and Language</subject><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2024</creationdate><recordtype>article</recordtype><sourceid>GOX</sourceid><recordid>eNpjYJA0NNAzsTA1NdBPLKrILNMzMjE01DMwNTYx5WRwCEpNLM7Py8xLVwjKTyotLslLLS5WyE9T8PHxLVYoyVdwTClLLSpOLMpMzFEIqSzITy9KLMjITAbyXIuK8ouKeRhY0xJzilN5oTQ3g7yba4izhy7YrviCoszcxKLKeJCd8WA7jQmrAAD_rjbP</recordid><startdate>20241108</startdate><enddate>20241108</enddate><creator>Gan, Esther</creator><creator>Zhao, Yiran</creator><creator>Cheng, Liying</creator><creator>Mao, Yancan</creator><creator>Goyal, Anirudh</creator><creator>Kawaguchi, Kenji</creator><creator>Kan, Min-Yen</creator><creator>Shieh, Michael</creator><scope>AKY</scope><scope>GOX</scope></search><sort><creationdate>20241108</creationdate><title>Reasoning Robustness of LLMs to Adversarial Typographical Errors</title><author>Gan, Esther ; Zhao, Yiran ; Cheng, Liying ; Mao, Yancan ; Goyal, Anirudh ; Kawaguchi, Kenji ; Kan, Min-Yen ; Shieh, Michael</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-arxiv_primary_2411_053453</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2024</creationdate><topic>Computer Science - Artificial Intelligence</topic><topic>Computer Science - Computation and Language</topic><toplevel>online_resources</toplevel><creatorcontrib>Gan, Esther</creatorcontrib><creatorcontrib>Zhao, Yiran</creatorcontrib><creatorcontrib>Cheng, Liying</creatorcontrib><creatorcontrib>Mao, Yancan</creatorcontrib><creatorcontrib>Goyal, Anirudh</creatorcontrib><creatorcontrib>Kawaguchi, Kenji</creatorcontrib><creatorcontrib>Kan, Min-Yen</creatorcontrib><creatorcontrib>Shieh, Michael</creatorcontrib><collection>arXiv Computer Science</collection><collection>arXiv.org</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext_linktorsrc</fulltext></delivery><addata><au>Gan, Esther</au><au>Zhao, Yiran</au><au>Cheng, Liying</au><au>Mao, Yancan</au><au>Goyal, Anirudh</au><au>Kawaguchi, Kenji</au><au>Kan, Min-Yen</au><au>Shieh, Michael</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>Reasoning Robustness of LLMs to Adversarial Typographical Errors</atitle><date>2024-11-08</date><risdate>2024</risdate><abstract>Large Language Models (LLMs) have demonstrated impressive capabilities in reasoning using Chain-of-Thought (CoT) prompting. However, CoT can be biased by users' instruction. In this work, we study the reasoning robustness of LLMs to typographical errors, which can naturally occur in users' queries. We design an Adversarial Typo Attack ($\texttt{ATA}$) algorithm that iteratively samples typos for words that are important to the query and selects the edit that is most likely to succeed in attacking. It shows that LLMs are sensitive to minimal adversarial typographical changes. Notably, with 1 character edit, Mistral-7B-Instruct's accuracy drops from 43.7% to 38.6% on GSM8K, while with 8 character edits the performance further drops to 19.2%. To extend our evaluation to larger and closed-source LLMs, we develop the $\texttt{R$^2$ATA}$ benchmark, which assesses models' $\underline{R}$easoning $\underline{R}$obustness to $\underline{\texttt{ATA}}$. It includes adversarial typographical questions derived from three widely used reasoning datasets-GSM8K, BBH, and MMLU-by applying $\texttt{ATA}$ to open-source LLMs. $\texttt{R$^2$ATA}$ demonstrates remarkable transferability and causes notable performance drops across multiple super large and closed-source LLMs.</abstract><doi>10.48550/arxiv.2411.05345</doi><oa>free_for_read</oa></addata></record>
fulltext fulltext_linktorsrc
identifier DOI: 10.48550/arxiv.2411.05345
ispartof
issn
language eng
recordid cdi_arxiv_primary_2411_05345
source arXiv.org
subjects Computer Science - Artificial Intelligence
Computer Science - Computation and Language
title Reasoning Robustness of LLMs to Adversarial Typographical Errors
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2024-12-17T07%3A49%3A42IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-arxiv_GOX&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=Reasoning%20Robustness%20of%20LLMs%20to%20Adversarial%20Typographical%20Errors&rft.au=Gan,%20Esther&rft.date=2024-11-08&rft_id=info:doi/10.48550/arxiv.2411.05345&rft_dat=%3Carxiv_GOX%3E2411_05345%3C/arxiv_GOX%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rfr_iscdi=true