<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="prins-etal-2026-model">
    <titleInfo>
        <title>Is my model perplexed for the right reason? Contrasting LLMs’ Benchmark Behavior with Token-Level Perplexity</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Zoë</namePart>
        <namePart type="family">Prins</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Samuele</namePart>
        <namePart type="family">Punzo</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Frank</namePart>
        <namePart type="family">Wildenburg</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Giovanni</namePart>
        <namePart type="family">Cinà</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Sandro</namePart>
        <namePart type="family">Pezzelle</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-10</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Hung-Yi</namePart>
            <namePart type="family">Lee</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Sunipa</namePart>
            <namePart type="family">Dev</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Jessy</namePart>
            <namePart type="family">Li</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Desmond</namePart>
            <namePart type="family">Elliott</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Budapest, Hungary</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-459-0</identifier>
    </relatedItem>
    <abstract>Standard evaluations of Large language models (LLMs) focus on task performance, offering limited insight into whether correct behavior reflects appropriate underlying mechanisms and risking confirmation bias. We introduce a simple, principled interpretability framework based on token-level perplexity to test whether models rely on linguistically relevant cues. By comparing perplexity distributions over minimal sentence pairs differing in one or a few ‘pivotal’ tokens, our method enables precise, hypothesis-driven analysis without relying on unstable feature-attribution techniques. Experiments on controlled linguistic benchmarks with several open-weight LLMs show that, while linguistically important tokens influence model behavior, they never fully explain perplexity shifts, revealing that models rely on heuristics other than the expected linguistic ones.</abstract>
    <identifier type="citekey">prins-etal-2026-model</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.110/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>2584</start>
            <end>2598</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Is my model perplexed for the right reason? Contrasting LLMs’ Benchmark Behavior with Token-Level Perplexity
%A Prins, Zoë
%A Punzo, Samuele
%A Wildenburg, Frank
%A Cinà, Giovanni
%A Pezzelle, Sandro
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F prins-etal-2026-model
%X Standard evaluations of Large language models (LLMs) focus on task performance, offering limited insight into whether correct behavior reflects appropriate underlying mechanisms and risking confirmation bias. We introduce a simple, principled interpretability framework based on token-level perplexity to test whether models rely on linguistically relevant cues. By comparing perplexity distributions over minimal sentence pairs differing in one or a few ‘pivotal’ tokens, our method enables precise, hypothesis-driven analysis without relying on unstable feature-attribution techniques. Experiments on controlled linguistic benchmarks with several open-weight LLMs show that, while linguistically important tokens influence model behavior, they never fully explain perplexity shifts, revealing that models rely on heuristics other than the expected linguistic ones.
%U https://aclanthology.org/2026.emnlp-main.110/
%P 2584-2598

Markdown (Informal)

[Is my model perplexed for the right reason? Contrasting LLMs’ Benchmark Behavior with Token-Level Perplexity](https://aclanthology.org/2026.emnlp-main.110/) (Prins et al., EMNLP 2026)

• Is my model perplexed for the right reason? Contrasting LLMs’ Benchmark Behavior with Token-Level Perplexity (Prins et al., EMNLP 2026)

ACL

• Zoë Prins, Samuele Punzo, Frank Wildenburg, Giovanni Cinà, and Sandro Pezzelle. 2026. Is my model perplexed for the right reason? Contrasting LLMs’ Benchmark Behavior with Token-Level Perplexity. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 2584–2598, Budapest, Hungary. Association for Computational Linguistics.