<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hegde-etal-2026-hallucinations">
    <titleInfo>
        <title>Where Hallucinations Live: A Cross-Architecture Circuit in VQ-Tokenized Vision-Language Models</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Shamanthak</namePart>
        <namePart type="family">Hegde</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Xiangrui</namePart>
        <namePart type="family">Liu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Maitreya</namePart>
        <namePart type="family">Patel</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Y</namePart>
        <namePart type="given">Z</namePart>
        <namePart type="family">Yang</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>Unified vision-language models (VLMs) that tokenize images through a vector-quantized (VQ) codebook routinely hallucinate objects on grounded yes/no benchmarks, yet existing decoding-time fixes treat this as generic miscalibration without an architectural account. Using activation patching across twenty-five models spanning eight LLM families, we identify an early-layer (L\ensuremath₀) attention routing circuit shared across VQ-tokenized VLMs and propose a three-gate diagnostic that distinguishes the models carrying it. The diagnostic isolates ten positive models (five natural unified-VQ VLMs across three LLM families and five induced variants) and rejects the remaining fifteen. A single-variable architectural swap (LLaVA-1.6 CLIP+MLP → VQ+Linear) installs the circuit, while a matched-compute MLP control does not, isolating vector quantization as the source of the pathological signal; the routing pathway that carries it is one that the backbone already provides. Against tuned baselines, DoLA wins on binary calibration, but only L\ensuremath₀ ablation reduces object hallucination in open-ended generation (CHAIRᵢ -31% relative; tuned DoLA and VCD leave it unchanged or worsen it). These results recast object hallucination in unified VQ VLMs as a property of architecture and pretraining, and yield a targeted intervention that mechanism-agnostic decoding cannot replicate.</abstract>
    <identifier type="citekey">hegde-etal-2026-hallucinations</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.2005/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>45671</start>
            <end>45689</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Where Hallucinations Live: A Cross-Architecture Circuit in VQ-Tokenized Vision-Language Models
%A Hegde, Shamanthak
%A Liu, Xiangrui
%A Patel, Maitreya
%A Yang, Y. Z.
%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 hegde-etal-2026-hallucinations
%X Unified vision-language models (VLMs) that tokenize images through a vector-quantized (VQ) codebook routinely hallucinate objects on grounded yes/no benchmarks, yet existing decoding-time fixes treat this as generic miscalibration without an architectural account. Using activation patching across twenty-five models spanning eight LLM families, we identify an early-layer (L\ensuremath₀) attention routing circuit shared across VQ-tokenized VLMs and propose a three-gate diagnostic that distinguishes the models carrying it. The diagnostic isolates ten positive models (five natural unified-VQ VLMs across three LLM families and five induced variants) and rejects the remaining fifteen. A single-variable architectural swap (LLaVA-1.6 CLIP+MLP → VQ+Linear) installs the circuit, while a matched-compute MLP control does not, isolating vector quantization as the source of the pathological signal; the routing pathway that carries it is one that the backbone already provides. Against tuned baselines, DoLA wins on binary calibration, but only L\ensuremath₀ ablation reduces object hallucination in open-ended generation (CHAIRᵢ -31% relative; tuned DoLA and VCD leave it unchanged or worsen it). These results recast object hallucination in unified VQ VLMs as a property of architecture and pretraining, and yield a targeted intervention that mechanism-agnostic decoding cannot replicate.
%U https://aclanthology.org/2026.emnlp-main.2005/
%P 45671-45689

Markdown (Informal)

[Where Hallucinations Live: A Cross-Architecture Circuit in VQ-Tokenized Vision-Language Models](https://aclanthology.org/2026.emnlp-main.2005/) (Hegde et al., EMNLP 2026)

• Where Hallucinations Live: A Cross-Architecture Circuit in VQ-Tokenized Vision-Language Models (Hegde et al., EMNLP 2026)

ACL

• Shamanthak Hegde, Xiangrui Liu, Maitreya Patel, and YZ Yang. 2026. Where Hallucinations Live: A Cross-Architecture Circuit in VQ-Tokenized Vision-Language Models. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 45671–45689, Budapest, Hungary. Association for Computational Linguistics.