@inproceedings{trirat-2026-modalities,
title = "More Modalities, More Problems: When Multimodal Agentic Debate Hurts Time-Series Reasoning",
author = "Trirat, Patara",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 2: Short Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-short.60/",
pages = "945--957",
ISBN = "979-8-89176-440-8",
abstract = "Multi-agent debate is often motivated by the premise that multiple agents, modalities, and rounds of deliberation improve reasoning by combining complementary evidence and correcting individual errors. We test this premise in multimodal time-series analysis using **MAD-TS**, a prompting-based framework with modality-specific agents, cross-modal judging, and synthesis. Contrary to expectation, MAD-TS does not reliably outperform simpler baselines. Across 20 tasks in finance, weather, healthcare, and science, it is uniquely best on 2 of 23 outcomes and significantly beats a strong single-agent multimodal baseline on only one task, while requiring about 22$\times$ more tokens and 12$\times$ more wall-clock time. Adding modalities, increasing debate depth, or changing judge rounds can help some tasks but harm others. Trace analysis attributes 85{\%} of errors to all three judges agreeing on the same wrong answer, a shared upstream bias that limits same-backbone multimodal debate as a general-purpose approach."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="trirat-2026-modalities">
<titleInfo>
<title>More Modalities, More Problems: When Multimodal Agentic Debate Hurts Time-Series Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Patara</namePart>
<namePart type="family">Trirat</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 2: Short Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-440-8</identifier>
</relatedItem>
<abstract>Multi-agent debate is often motivated by the premise that multiple agents, modalities, and rounds of deliberation improve reasoning by combining complementary evidence and correcting individual errors. We test this premise in multimodal time-series analysis using **MAD-TS**, a prompting-based framework with modality-specific agents, cross-modal judging, and synthesis. Contrary to expectation, MAD-TS does not reliably outperform simpler baselines. Across 20 tasks in finance, weather, healthcare, and science, it is uniquely best on 2 of 23 outcomes and significantly beats a strong single-agent multimodal baseline on only one task, while requiring about 22\times more tokens and 12\times more wall-clock time. Adding modalities, increasing debate depth, or changing judge rounds can help some tasks but harm others. Trace analysis attributes 85% of errors to all three judges agreeing on the same wrong answer, a shared upstream bias that limits same-backbone multimodal debate as a general-purpose approach.</abstract>
<identifier type="citekey">trirat-2026-modalities</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-short.60/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>945</start>
<end>957</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T More Modalities, More Problems: When Multimodal Agentic Debate Hurts Time-Series Reasoning
%A Trirat, Patara
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 2: Short Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-440-8
%F trirat-2026-modalities
%X Multi-agent debate is often motivated by the premise that multiple agents, modalities, and rounds of deliberation improve reasoning by combining complementary evidence and correcting individual errors. We test this premise in multimodal time-series analysis using **MAD-TS**, a prompting-based framework with modality-specific agents, cross-modal judging, and synthesis. Contrary to expectation, MAD-TS does not reliably outperform simpler baselines. Across 20 tasks in finance, weather, healthcare, and science, it is uniquely best on 2 of 23 outcomes and significantly beats a strong single-agent multimodal baseline on only one task, while requiring about 22\times more tokens and 12\times more wall-clock time. Adding modalities, increasing debate depth, or changing judge rounds can help some tasks but harm others. Trace analysis attributes 85% of errors to all three judges agreeing on the same wrong answer, a shared upstream bias that limits same-backbone multimodal debate as a general-purpose approach.
%U https://aclanthology.org/2026.aacl-short.60/
%P 945-957
Markdown (Informal)
[More Modalities, More Problems: When Multimodal Agentic Debate Hurts Time-Series Reasoning](https://aclanthology.org/2026.aacl-short.60/) (Trirat, AACL-IJCNLP 2026)
ACL