<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hejabi-etal-2026-flip">
    <titleInfo>
        <title>Flip-Flop Consistency: Unsupervised Training for Robustness to Prompt Perturbations in LLMs</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Parsa</namePart>
        <namePart type="family">Hejabi</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Elnaz</namePart>
        <namePart type="family">Rahmati</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Alireza</namePart>
        <namePart type="family">Salkhordeh Ziabari</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Morteza</namePart>
        <namePart type="family">Dehghani</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-07</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Maria</namePart>
            <namePart type="family">Liakata</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Viviane</namePart>
            <namePart type="given">P</namePart>
            <namePart type="family">Moreira</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Jiajun</namePart>
            <namePart type="family">Zhang</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">David</namePart>
            <namePart type="family">Jurgens</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">San Diego, California, United States</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-390-6</identifier>
    </relatedItem>
    <abstract>Large Language Models (LLMs) often produce inconsistent answers when faced with different phrasings of the same prompt. In this paper, we propose Flip-Flop Consistency (F²C), an unsupervised training method that improves robustness to such perturbations. F²C is composed of two key components. The first, Consensus Cross-Entropy (CCE), uses a majority vote across prompt variations to create a hard pseudo-label. The second is a representation alignment loss that pulls lower-confidence and non-majority predictors toward the consensus established by high-confidence, majority-voting variations. We evaluate our method on 11 datasets spanning four NLP tasks, with 4–15 prompt variations per dataset. On average, F²C raises observed agreement by 11.62%, improves mean F₁ by 8.94%, and reduces performance variance across formats by 3.29%. In out-of-domain evaluations, F²C generalizes effectively, increasing øverlineF₁ and agreement while decreasing variance across most source-target pairs. Finally, when trained on only a subset of prompt perturbations and evaluated on held-out formats, F²C consistently improves both performance and agreement while reducing variance. These findings highlight F²C as an effective unsupervised method for enhancing LLM consistency, performance, and generalization under prompt perturbations.</abstract>
    <identifier type="citekey">hejabi-etal-2026-flip</identifier>
    <location>
        <url>https://aclanthology.org/2026.acl-long.71/</url>
    </location>
    <part>
        <date>2026-07</date>
        <extent unit="page">
            <start>1571</start>
            <end>1587</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Flip-Flop Consistency: Unsupervised Training for Robustness to Prompt Perturbations in LLMs
%A Hejabi, Parsa
%A Rahmati, Elnaz
%A Salkhordeh Ziabari, Alireza
%A Dehghani, Morteza
%Y Liakata, Maria
%Y Moreira, Viviane P.
%Y Zhang, Jiajun
%Y Jurgens, David
%S Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California, United States
%@ 979-8-89176-390-6
%F hejabi-etal-2026-flip
%X Large Language Models (LLMs) often produce inconsistent answers when faced with different phrasings of the same prompt. In this paper, we propose Flip-Flop Consistency (F²C), an unsupervised training method that improves robustness to such perturbations. F²C is composed of two key components. The first, Consensus Cross-Entropy (CCE), uses a majority vote across prompt variations to create a hard pseudo-label. The second is a representation alignment loss that pulls lower-confidence and non-majority predictors toward the consensus established by high-confidence, majority-voting variations. We evaluate our method on 11 datasets spanning four NLP tasks, with 4–15 prompt variations per dataset. On average, F²C raises observed agreement by 11.62%, improves mean F₁ by 8.94%, and reduces performance variance across formats by 3.29%. In out-of-domain evaluations, F²C generalizes effectively, increasing øverlineF₁ and agreement while decreasing variance across most source-target pairs. Finally, when trained on only a subset of prompt perturbations and evaluated on held-out formats, F²C consistently improves both performance and agreement while reducing variance. These findings highlight F²C as an effective unsupervised method for enhancing LLM consistency, performance, and generalization under prompt perturbations.
%U https://aclanthology.org/2026.acl-long.71/
%P 1571-1587

Markdown (Informal)

[Flip-Flop Consistency: Unsupervised Training for Robustness to Prompt Perturbations in LLMs](https://aclanthology.org/2026.acl-long.71/) (Hejabi et al., ACL 2026)

• Flip-Flop Consistency: Unsupervised Training for Robustness to Prompt Perturbations in LLMs (Hejabi et al., ACL 2026)

ACL

• Parsa Hejabi, Elnaz Rahmati, Alireza Salkhordeh Ziabari, and Morteza Dehghani. 2026. Flip-Flop Consistency: Unsupervised Training for Robustness to Prompt Perturbations in LLMs. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1571–1587, San Diego, California, United States. Association for Computational Linguistics.