@inproceedings{zheng-etal-2026-well,
title = "How Well Can {M}odern {LLM}s Act as Agent Cores in Radiology Environments?",
author = "Zheng, Qiaoyu and
Wu, Chaoyi and
Zhao, Weike and
Qiu, Pengcheng and
Dai, Lisong and
Zhang, Ya and
Wang, Yanfeng and
Xie, Weidi",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.1000/",
pages = "23010--23040",
ISBN = "979-8-89176-459-0",
abstract = "Radiology, with its heterogeneous modalities, anatomies, and evolving protocols, is a natural yet high-stakes testbed for agentic AI. As LLMs grow more capable and tool ecosystems such as MCP and Agent Skills more complex, their capability boundaries in radiology remain unclear. We introduce \textbf{RadA-BenchPlat}, a two-layer benchmark spanning idealized reasoning and real-world execution: a \textit{synthetic layer} with 2.2k clinician-verified records, 24.2k QA pairs, and 10 tool categories under diverse availability settings; and a \textit{real-environment layer} pairing 165 2D/3D cases with executable tools for grounding, diagnosis, and report generation. We find that high completion rates under idealized settings do not carry over to real environments, where the best agents trail their upper bound by $\sim$15{\%} and drop to 0.327 on long-chain tasks. Prompting strategies (few-shot exemplars, multi-agent coordination) and AutoTB (on-the-fly synthesis of missing tools) mitigate execution drift and breakdowns in difficult scenarios{---}lifting execution success to 94.5{\%}{---}yet remain not fully reliable. Overall, our findings suggest that radiology agents are no longer limited to conceptual demonstrations, but are beginning to show potential for real-world applications."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zheng-etal-2026-well">
<titleInfo>
<title>How Well Can Modern LLMs Act as Agent Cores in Radiology Environments?</title>
</titleInfo>
<name type="personal">
<namePart type="given">Qiaoyu</namePart>
<namePart type="family">Zheng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chaoyi</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Weike</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pengcheng</namePart>
<namePart type="family">Qiu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lisong</namePart>
<namePart type="family">Dai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ya</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yanfeng</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Weidi</namePart>
<namePart type="family">Xie</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>Radiology, with its heterogeneous modalities, anatomies, and evolving protocols, is a natural yet high-stakes testbed for agentic AI. As LLMs grow more capable and tool ecosystems such as MCP and Agent Skills more complex, their capability boundaries in radiology remain unclear. We introduce RadA-BenchPlat, a two-layer benchmark spanning idealized reasoning and real-world execution: a synthetic layer with 2.2k clinician-verified records, 24.2k QA pairs, and 10 tool categories under diverse availability settings; and a real-environment layer pairing 165 2D/3D cases with executable tools for grounding, diagnosis, and report generation. We find that high completion rates under idealized settings do not carry over to real environments, where the best agents trail their upper bound by \sim15% and drop to 0.327 on long-chain tasks. Prompting strategies (few-shot exemplars, multi-agent coordination) and AutoTB (on-the-fly synthesis of missing tools) mitigate execution drift and breakdowns in difficult scenarios—lifting execution success to 94.5%—yet remain not fully reliable. Overall, our findings suggest that radiology agents are no longer limited to conceptual demonstrations, but are beginning to show potential for real-world applications.</abstract>
<identifier type="citekey">zheng-etal-2026-well</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1000/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>23010</start>
<end>23040</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T How Well Can Modern LLMs Act as Agent Cores in Radiology Environments?
%A Zheng, Qiaoyu
%A Wu, Chaoyi
%A Zhao, Weike
%A Qiu, Pengcheng
%A Dai, Lisong
%A Zhang, Ya
%A Wang, Yanfeng
%A Xie, Weidi
%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 zheng-etal-2026-well
%X Radiology, with its heterogeneous modalities, anatomies, and evolving protocols, is a natural yet high-stakes testbed for agentic AI. As LLMs grow more capable and tool ecosystems such as MCP and Agent Skills more complex, their capability boundaries in radiology remain unclear. We introduce RadA-BenchPlat, a two-layer benchmark spanning idealized reasoning and real-world execution: a synthetic layer with 2.2k clinician-verified records, 24.2k QA pairs, and 10 tool categories under diverse availability settings; and a real-environment layer pairing 165 2D/3D cases with executable tools for grounding, diagnosis, and report generation. We find that high completion rates under idealized settings do not carry over to real environments, where the best agents trail their upper bound by \sim15% and drop to 0.327 on long-chain tasks. Prompting strategies (few-shot exemplars, multi-agent coordination) and AutoTB (on-the-fly synthesis of missing tools) mitigate execution drift and breakdowns in difficult scenarios—lifting execution success to 94.5%—yet remain not fully reliable. Overall, our findings suggest that radiology agents are no longer limited to conceptual demonstrations, but are beginning to show potential for real-world applications.
%U https://aclanthology.org/2026.emnlp-main.1000/
%P 23010-23040
Markdown (Informal)
[How Well Can Modern LLMs Act as Agent Cores in Radiology Environments?](https://aclanthology.org/2026.emnlp-main.1000/) (Zheng et al., EMNLP 2026)
ACL
- Qiaoyu Zheng, Chaoyi Wu, Weike Zhao, Pengcheng Qiu, Lisong Dai, Ya Zhang, Yanfeng Wang, and Weidi Xie. 2026. How Well Can Modern LLMs Act as Agent Cores in Radiology Environments?. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 23010–23040, Budapest, Hungary. Association for Computational Linguistics.