@inproceedings{hu-etal-2026-seeing,
title = "Seeing Failure In Advance: Pre-Generation Failure Diagnosis for Multi-Modal Retrieval-Augmented-Generation",
author = "Hu, Chan-Wei and
Xing, Shuo and
Rossi, Ryan A. and
Tu, Zhengzhong",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.616/",
doi = "10.18653/2026.findings-emnlp.616",
pages = "12959--12977",
ISBN = "979-8-89176-462-0",
abstract = "Multimodal retrieval-augmented generation (MMRAG) systems can fail even when relevant evidence is retrieved, exposing a gap between retrieval and generation. We formulate MMRAG failure diagnosis as an online single-fire decision problem over prefill hidden states, which is distinct from static probing of model internals or post-hoc self-correction. Across InfoSeek, E-VQA, and VISA, we show that hidden states encode both failure occurrence and failure type, with failure detection reaching around 90{\%} accuracy and recall on InfoSeek and E-VQA. We further introduce a Fire-Then-Type prober that separates failure detection from failure typing, improving typed diagnosis over direct classification. Finally, our method matches the response accuracy of final-layer repair on InfoSeek and E-VQA while diagnosing before answer decoding. The available prefill depth varies substantially across models, from 4.6{\%} to 24.3{\%}, and is limited for Qwen3-VL and InternVL on VISA."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hu-etal-2026-seeing">
<titleInfo>
<title>Seeing Failure In Advance: Pre-Generation Failure Diagnosis for Multi-Modal Retrieval-Augmented-Generation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Chan-Wei</namePart>
<namePart type="family">Hu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shuo</namePart>
<namePart type="family">Xing</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ryan</namePart>
<namePart type="given">A</namePart>
<namePart type="family">Rossi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhengzhong</namePart>
<namePart type="family">Tu</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>Multimodal retrieval-augmented generation (MMRAG) systems can fail even when relevant evidence is retrieved, exposing a gap between retrieval and generation. We formulate MMRAG failure diagnosis as an online single-fire decision problem over prefill hidden states, which is distinct from static probing of model internals or post-hoc self-correction. Across InfoSeek, E-VQA, and VISA, we show that hidden states encode both failure occurrence and failure type, with failure detection reaching around 90% accuracy and recall on InfoSeek and E-VQA. We further introduce a Fire-Then-Type prober that separates failure detection from failure typing, improving typed diagnosis over direct classification. Finally, our method matches the response accuracy of final-layer repair on InfoSeek and E-VQA while diagnosing before answer decoding. The available prefill depth varies substantially across models, from 4.6% to 24.3%, and is limited for Qwen3-VL and InternVL on VISA.</abstract>
<identifier type="citekey">hu-etal-2026-seeing</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.616</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.616/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>12959</start>
<end>12977</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Seeing Failure In Advance: Pre-Generation Failure Diagnosis for Multi-Modal Retrieval-Augmented-Generation
%A Hu, Chan-Wei
%A Xing, Shuo
%A Rossi, Ryan A.
%A Tu, Zhengzhong
%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 hu-etal-2026-seeing
%X Multimodal retrieval-augmented generation (MMRAG) systems can fail even when relevant evidence is retrieved, exposing a gap between retrieval and generation. We formulate MMRAG failure diagnosis as an online single-fire decision problem over prefill hidden states, which is distinct from static probing of model internals or post-hoc self-correction. Across InfoSeek, E-VQA, and VISA, we show that hidden states encode both failure occurrence and failure type, with failure detection reaching around 90% accuracy and recall on InfoSeek and E-VQA. We further introduce a Fire-Then-Type prober that separates failure detection from failure typing, improving typed diagnosis over direct classification. Finally, our method matches the response accuracy of final-layer repair on InfoSeek and E-VQA while diagnosing before answer decoding. The available prefill depth varies substantially across models, from 4.6% to 24.3%, and is limited for Qwen3-VL and InternVL on VISA.
%R 10.18653/2026.findings-emnlp.616
%U https://aclanthology.org/2026.findings-emnlp.616/
%U https://doi.org/10.18653/2026.findings-emnlp.616
%P 12959-12977
Markdown (Informal)
[Seeing Failure In Advance: Pre-Generation Failure Diagnosis for Multi-Modal Retrieval-Augmented-Generation](https://aclanthology.org/2026.findings-emnlp.616/) (Hu et al., Findings 2026)
ACL