<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="bansal-etal-2026-vibe">
    <titleInfo>
        <title>Can Vibe Coders Really Pass the Vibe Check? An Empirical Analysis of Fault-Targeted Reasoning in Frontier LLMs</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Srijan</namePart>
        <namePart type="family">Bansal</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Fangkai</namePart>
        <namePart type="family">Jiao</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yilun</namePart>
        <namePart type="family">Zhou</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Austin</namePart>
        <namePart type="family">Xu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Shafiq</namePart>
        <namePart type="family">Joty</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Semih</namePart>
        <namePart type="family">Yavuz</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>Findings of the Association for Computational Linguistics: EMNLP 2026</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">979-8-89176-462-0</identifier>
    </relatedItem>
    <abstract>We conduct a controlled study isolating reasoning in autonomous debugging by holding scaffolding and tool access fixed. We decompose fault-targeted reasoning into test generation and repair using LLM-generated solutions that pass most tests but fail on edge cases. Across 12 frontier and open-source models, we identify three limitations. First, despite 87.7% input validity, only 68.3% of inputs are discriminative, identifying fault-hypothesis formation as the bottleneck: finding fault-triggering inputs costs 19.4 points versus 2.7 points for output prediction. Second, diagnostic feedback has mixed effects on repair: tests generated internally within the repair response reduce success by 5.6 points relative to a bug label alone, while separately generated same-model tests differ little on average, but repair succeeds more often with discriminative than non-discriminative tests (66.8% versus 25.2%). Third, capabilities do not consistently compose: success across bug detection, test construction, and repair does not consistently occur on the same instances, with the largest losses in forming a fault-triggering input and converting a discriminative test into a correct repair. Together, these findings identify both transitions as key debugging limits.</abstract>
    <identifier type="citekey">bansal-etal-2026-vibe</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-emnlp.2389/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>51358</start>
            <end>51371</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Can Vibe Coders Really Pass the Vibe Check? An Empirical Analysis of Fault-Targeted Reasoning in Frontier LLMs
%A Bansal, Srijan
%A Jiao, Fangkai
%A Zhou, Yilun
%A Xu, Austin
%A Joty, Shafiq
%A Yavuz, Semih
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F bansal-etal-2026-vibe
%X We conduct a controlled study isolating reasoning in autonomous debugging by holding scaffolding and tool access fixed. We decompose fault-targeted reasoning into test generation and repair using LLM-generated solutions that pass most tests but fail on edge cases. Across 12 frontier and open-source models, we identify three limitations. First, despite 87.7% input validity, only 68.3% of inputs are discriminative, identifying fault-hypothesis formation as the bottleneck: finding fault-triggering inputs costs 19.4 points versus 2.7 points for output prediction. Second, diagnostic feedback has mixed effects on repair: tests generated internally within the repair response reduce success by 5.6 points relative to a bug label alone, while separately generated same-model tests differ little on average, but repair succeeds more often with discriminative than non-discriminative tests (66.8% versus 25.2%). Third, capabilities do not consistently compose: success across bug detection, test construction, and repair does not consistently occur on the same instances, with the largest losses in forming a fault-triggering input and converting a discriminative test into a correct repair. Together, these findings identify both transitions as key debugging limits.
%U https://aclanthology.org/2026.findings-emnlp.2389/
%P 51358-51371

Markdown (Informal)

[Can Vibe Coders Really Pass the Vibe Check? An Empirical Analysis of Fault-Targeted Reasoning in Frontier LLMs](https://aclanthology.org/2026.findings-emnlp.2389/) (Bansal et al., Findings 2026)

• Can Vibe Coders Really Pass the Vibe Check? An Empirical Analysis of Fault-Targeted Reasoning in Frontier LLMs (Bansal et al., Findings 2026)

ACL

• Srijan Bansal, Fangkai Jiao, Yilun Zhou, Austin Xu, Shafiq Joty, and Semih Yavuz. 2026. Can Vibe Coders Really Pass the Vibe Check? An Empirical Analysis of Fault-Targeted Reasoning in Frontier LLMs. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 51358–51371, Budapest, Hungary. Association for Computational Linguistics.