<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="jalan-etal-2026-medbench">
    <titleInfo>
        <title>MedBench: Deliberative Evaluation of Medical Language Models</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Pratik</namePart>
        <namePart type="family">Jalan</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Mukul</namePart>
        <namePart type="family">Joshi</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Akhilesh</namePart>
        <namePart type="family">Magotra</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Kshitij</namePart>
        <namePart type="family">Jadhav</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-07</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>BioNLP 2026</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Dina</namePart>
            <namePart type="family">Demner-Fushman</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Sophia</namePart>
            <namePart type="family">Ananiadou</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Kirk</namePart>
            <namePart type="family">Roberts</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Junichi</namePart>
            <namePart type="family">Tsujii</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">San Diego, California</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-434-7</identifier>
    </relatedItem>
    <abstract>We introduce MedBench, a benchmark for evaluating medical language models as deliberating agents rather than isolated predictors. MedBench evaluates eight models (4B?32B) on 19,625 questions from six medical QA datasets using Consensus-Aware Model Panel (CAMP), a two-tier protocol in which five 4B?8B models answer independently, revise after observing peer reasoning, and escalate persistent disagreements to larger 20B?32B models. Compared with zero-shot, few-shot, and chain-of-thought baselines, CAMP shows that deliberation is not uniformly accuracy-improving, but reveals interaction-driven behaviors hidden by single-model evaluation. On PubMedQA without external context, the 4B?8B panel outperforms the evaluated 20B?32B individual zero-shot models (54.1% vs. 33.9%), and achieves the best evaluated result with context (75.7%), suggesting that structured interaction can sometimes complement scale. Across five datasets, initial inter-model agreement is positively associated with correctness and serves as a useful difficulty signal. However, on MedXpertQA, unanimous agreement yields only 6.6% accuracy despite 14.4% overall accuracy, suggesting correlated ignorance, where shared biases make consensus misleading. Error analysis shows that most failures are debate-insufficient cases, where incorrect majorities persist despite interaction (93?97%), while debate-harmful cases account for 3?7%. MedBench positions deliberative evaluation as a complement to accuracy-centric benchmarking, measuring when model interaction corrects errors, reinforces shared mistakes, or signals the need for stronger evidence and human review.</abstract>
    <identifier type="citekey">jalan-etal-2026-medbench</identifier>
    <location>
        <url>https://aclanthology.org/2026.bionlp-1.79/</url>
    </location>
    <part>
        <date>2026-07</date>
        <extent unit="page">
            <start>981</start>
            <end>991</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T MedBench: Deliberative Evaluation of Medical Language Models
%A Jalan, Pratik
%A Joshi, Mukul
%A Magotra, Akhilesh
%A Jadhav, Kshitij
%Y Demner-Fushman, Dina
%Y Ananiadou, Sophia
%Y Roberts, Kirk
%Y Tsujii, Junichi
%S BioNLP 2026
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California
%@ 979-8-89176-434-7
%F jalan-etal-2026-medbench
%X We introduce MedBench, a benchmark for evaluating medical language models as deliberating agents rather than isolated predictors. MedBench evaluates eight models (4B?32B) on 19,625 questions from six medical QA datasets using Consensus-Aware Model Panel (CAMP), a two-tier protocol in which five 4B?8B models answer independently, revise after observing peer reasoning, and escalate persistent disagreements to larger 20B?32B models. Compared with zero-shot, few-shot, and chain-of-thought baselines, CAMP shows that deliberation is not uniformly accuracy-improving, but reveals interaction-driven behaviors hidden by single-model evaluation. On PubMedQA without external context, the 4B?8B panel outperforms the evaluated 20B?32B individual zero-shot models (54.1% vs. 33.9%), and achieves the best evaluated result with context (75.7%), suggesting that structured interaction can sometimes complement scale. Across five datasets, initial inter-model agreement is positively associated with correctness and serves as a useful difficulty signal. However, on MedXpertQA, unanimous agreement yields only 6.6% accuracy despite 14.4% overall accuracy, suggesting correlated ignorance, where shared biases make consensus misleading. Error analysis shows that most failures are debate-insufficient cases, where incorrect majorities persist despite interaction (93?97%), while debate-harmful cases account for 3?7%. MedBench positions deliberative evaluation as a complement to accuracy-centric benchmarking, measuring when model interaction corrects errors, reinforces shared mistakes, or signals the need for stronger evidence and human review.
%U https://aclanthology.org/2026.bionlp-1.79/
%P 981-991

Markdown (Informal)

[MedBench: Deliberative Evaluation of Medical Language Models](https://aclanthology.org/2026.bionlp-1.79/) (Jalan et al., BioNLP 2026)

• MedBench: Deliberative Evaluation of Medical Language Models (Jalan et al., BioNLP 2026)

ACL

• Pratik Jalan, Mukul Joshi, Akhilesh Magotra, and Kshitij Jadhav. 2026. MedBench: Deliberative Evaluation of Medical Language Models. In BioNLP 2026, pages 981–991, San Diego, California. Association for Computational Linguistics.