@inproceedings{kulkarni-etal-2026-answer,
title = "The Same Answer Every Time: Auditing Small Medical {LLM} Agents",
author = "Kulkarni, Abhineet Rajendra and
Kuber, Pranav Madhav and
Bihani, Anurag",
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 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.175/",
pages = "3915--3933",
ISBN = "979-8-89176-439-2",
abstract = "We investigate how small open-weight language models make decisions in medical tasks and whether their failures also occur in other settings. We evaluated nine models on wearable fatigue and intensive-care tasks, where they chose among six recommendations, including continued monitoring. We then extended testing to a non-medical tool router and subsets of the Berkeley Function-Calling Leaderboard and JSONSchemaBench. We introduce the Minimum Viable Audit (MVA) to examine which actions models choose, whether their responses contain usable decisions, and whether they copy prompt content. Comparisons across tasks help establish where these behaviors recur. Under the baseline intensive-care prompt, MedGemma chose continued monitoring in all 450 attempts, achieving 33.3{\%} accuracy. On the router, it used all six tools and answered 144 of 150 requests correctly. Meditron-7B produced unusable responses and copied prompt content across intensive-care monitoring and the non-medical evaluations at both four-bit and eight-bit quantization. MedGemma{'}s results motivate checking the medical instructions. Meditron{'}s recurring failures warrant checking the model configuration and how outputs are generated. Differences between tasks and mismatches between medical prompts and scoring rules limit conclusions about the causes of these failures. We recommend reporting selected actions, invalid and copied responses, and results for each prompt condition alongside accuracy."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kulkarni-etal-2026-answer">
<titleInfo>
<title>The Same Answer Every Time: Auditing Small Medical LLM Agents</title>
</titleInfo>
<name type="personal">
<namePart type="given">Abhineet</namePart>
<namePart type="given">Rajendra</namePart>
<namePart type="family">Kulkarni</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pranav</namePart>
<namePart type="given">Madhav</namePart>
<namePart type="family">Kuber</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anurag</namePart>
<namePart type="family">Bihani</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 1: Long 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">979-8-89176-439-2</identifier>
</relatedItem>
<abstract>We investigate how small open-weight language models make decisions in medical tasks and whether their failures also occur in other settings. We evaluated nine models on wearable fatigue and intensive-care tasks, where they chose among six recommendations, including continued monitoring. We then extended testing to a non-medical tool router and subsets of the Berkeley Function-Calling Leaderboard and JSONSchemaBench. We introduce the Minimum Viable Audit (MVA) to examine which actions models choose, whether their responses contain usable decisions, and whether they copy prompt content. Comparisons across tasks help establish where these behaviors recur. Under the baseline intensive-care prompt, MedGemma chose continued monitoring in all 450 attempts, achieving 33.3% accuracy. On the router, it used all six tools and answered 144 of 150 requests correctly. Meditron-7B produced unusable responses and copied prompt content across intensive-care monitoring and the non-medical evaluations at both four-bit and eight-bit quantization. MedGemma’s results motivate checking the medical instructions. Meditron’s recurring failures warrant checking the model configuration and how outputs are generated. Differences between tasks and mismatches between medical prompts and scoring rules limit conclusions about the causes of these failures. We recommend reporting selected actions, invalid and copied responses, and results for each prompt condition alongside accuracy.</abstract>
<identifier type="citekey">kulkarni-etal-2026-answer</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.175/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>3915</start>
<end>3933</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T The Same Answer Every Time: Auditing Small Medical LLM Agents
%A Kulkarni, Abhineet Rajendra
%A Kuber, Pranav Madhav
%A Bihani, Anurag
%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 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F kulkarni-etal-2026-answer
%X We investigate how small open-weight language models make decisions in medical tasks and whether their failures also occur in other settings. We evaluated nine models on wearable fatigue and intensive-care tasks, where they chose among six recommendations, including continued monitoring. We then extended testing to a non-medical tool router and subsets of the Berkeley Function-Calling Leaderboard and JSONSchemaBench. We introduce the Minimum Viable Audit (MVA) to examine which actions models choose, whether their responses contain usable decisions, and whether they copy prompt content. Comparisons across tasks help establish where these behaviors recur. Under the baseline intensive-care prompt, MedGemma chose continued monitoring in all 450 attempts, achieving 33.3% accuracy. On the router, it used all six tools and answered 144 of 150 requests correctly. Meditron-7B produced unusable responses and copied prompt content across intensive-care monitoring and the non-medical evaluations at both four-bit and eight-bit quantization. MedGemma’s results motivate checking the medical instructions. Meditron’s recurring failures warrant checking the model configuration and how outputs are generated. Differences between tasks and mismatches between medical prompts and scoring rules limit conclusions about the causes of these failures. We recommend reporting selected actions, invalid and copied responses, and results for each prompt condition alongside accuracy.
%U https://aclanthology.org/2026.aacl-long.175/
%P 3915-3933
Markdown (Informal)
[The Same Answer Every Time: Auditing Small Medical LLM Agents](https://aclanthology.org/2026.aacl-long.175/) (Kulkarni et al., AACL-IJCNLP 2026)
ACL
- Abhineet Rajendra Kulkarni, Pranav Madhav Kuber, and Anurag Bihani. 2026. The Same Answer Every Time: Auditing Small Medical LLM Agents. In 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 1: Long Papers), pages 3915–3933, Hengqin, China. Association for Computational Linguistics.