<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="purbey-etal-2026-near">
    <titleInfo>
        <title>Near-Miss Context Contamination and Language-Asymmetric Evidence Arbitration in LLMs</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Jebish</namePart>
        <namePart type="family">Purbey</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Sanjeeb</namePart>
        <namePart type="given">Prasad</namePart>
        <namePart type="family">Panday</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Barnnita</namePart>
        <namePart type="family">Shrestha</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Bikash</namePart>
        <namePart type="family">Balami</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Aakash</namePart>
        <namePart type="family">Pandey</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Bibek</namePart>
        <namePart type="family">Khanal</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>Retrieval-augmented generation assumes that language models can use relevant context while rejecting semantically related but non-evidential context. Further, multilingual retrieval adds a second assumption that when evidence appears in different languages, models arbitrate by factual content instead of language cues. We test both these assumptions as generator-side failures to understand how models bind, compare, and trust evidence across languages. We first study near-miss contamination where we add true, relation-matched context about a nearby entity that does not answer the target question. Across Mintaka and PopQA, this context raises distractor adoption before any direct contradiction is introduced, and the effect persists across every tested model and language.\par We then study language evidence arbitration where English and local-language evidence disagree. We find that language preference is model-family dependent. GPT-5.4-mini leans toward English across all four languages, while DeepSeek models lean toward local-language evidence for all tested languages. We also run a label-only control where both snippets are written in English and only the source label varies and find that evidence labeled as English is over-trusted, with 95% bootstrap confidence intervals excluding zero across all 16 model-language pairs. Our results thus show that context use is shaped by entity proximity, language preference, and source-language metadata.</abstract>
    <identifier type="citekey">purbey-etal-2026-near</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.2392/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>54388</start>
            <end>54404</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Near-Miss Context Contamination and Language-Asymmetric Evidence Arbitration in LLMs
%A Purbey, Jebish
%A Panday, Sanjeeb Prasad
%A Shrestha, Barnnita
%A Balami, Bikash
%A Pandey, Aakash
%A Khanal, Bibek
%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 purbey-etal-2026-near
%X Retrieval-augmented generation assumes that language models can use relevant context while rejecting semantically related but non-evidential context. Further, multilingual retrieval adds a second assumption that when evidence appears in different languages, models arbitrate by factual content instead of language cues. We test both these assumptions as generator-side failures to understand how models bind, compare, and trust evidence across languages. We first study near-miss contamination where we add true, relation-matched context about a nearby entity that does not answer the target question. Across Mintaka and PopQA, this context raises distractor adoption before any direct contradiction is introduced, and the effect persists across every tested model and language.\par We then study language evidence arbitration where English and local-language evidence disagree. We find that language preference is model-family dependent. GPT-5.4-mini leans toward English across all four languages, while DeepSeek models lean toward local-language evidence for all tested languages. We also run a label-only control where both snippets are written in English and only the source label varies and find that evidence labeled as English is over-trusted, with 95% bootstrap confidence intervals excluding zero across all 16 model-language pairs. Our results thus show that context use is shaped by entity proximity, language preference, and source-language metadata.
%U https://aclanthology.org/2026.emnlp-main.2392/
%P 54388-54404

Markdown (Informal)

[Near-Miss Context Contamination and Language-Asymmetric Evidence Arbitration in LLMs](https://aclanthology.org/2026.emnlp-main.2392/) (Purbey et al., EMNLP 2026)

• Near-Miss Context Contamination and Language-Asymmetric Evidence Arbitration in LLMs (Purbey et al., EMNLP 2026)

ACL

• Jebish Purbey, Sanjeeb Prasad Panday, Barnnita Shrestha, Bikash Balami, Aakash Pandey, and Bibek Khanal. 2026. Near-Miss Context Contamination and Language-Asymmetric Evidence Arbitration in LLMs. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 54388–54404, Budapest, Hungary. Association for Computational Linguistics.