<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wu-etal-2026-balancedbio">
    <titleInfo>
        <title>BalancedBio: Mitigating the Alignment Tax in Biomedical LLMs via Gradient Orthogonality and GRPO</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Wentao</namePart>
        <namePart type="family">Wu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Hanmeng</namePart>
        <namePart type="family">Zhong</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Linqing</namePart>
        <namePart type="family">Chen</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Peng</namePart>
        <namePart type="family">Zhou</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>Aligning Large Language Models (LLMs) for specialized domains presents a fundamental optimization challenge: the “alignment tax.” In biomedicine, this manifests as a conflict between the need for rigorous, encyclopedic factual accuracy and the requirement for flexible, user-friendly instruction following. Existing methods, primarily relying on Supervised Fine-Tuning (SFT) or standard Reinforcement Learning from Human Feedback (RLHF), often fail to navigate this Pareto frontier, resulting in models that are either knowledgeable but rigid, or chatty but hallucination-prone. In this paper, we propose BalancedBio, a novel alignment framework that explicitly models and optimizes conflicting objectives. While utilizing standard high-quality instruction data, our contribution lies in the algorithmic innovation: (1) We introduce Capability-Aware Group Relative Policy Optimization (GRPO), which eliminates the need for a value network and reduces gradient variance in high-entropy reasoning tasks; (2) We propose a Dynamic Hybrid Reward Mechanism that adaptively balances domain correctness, reasoning validity, and format compliance during training; (3) We provide a theoretical analysis demonstrating how our method enforces gradient orthogonality to mitigate catastrophic forgetting. Extensive experiments on BIOMED-MMLU, MedQA, and IFEval show that BalancedBio-7B achieves state-of-the-art performance, surpassing Med-PaLM-8B in domain tasks while maintaining robust general instruction-following capabilities. We will release our model and partial data.</abstract>
    <identifier type="citekey">wu-etal-2026-balancedbio</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.2556/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>58101</start>
            <end>58115</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T BalancedBio: Mitigating the Alignment Tax in Biomedical LLMs via Gradient Orthogonality and GRPO
%A Wu, Wentao
%A Zhong, Hanmeng
%A Chen, Linqing
%A Zhou, Peng
%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 wu-etal-2026-balancedbio
%X Aligning Large Language Models (LLMs) for specialized domains presents a fundamental optimization challenge: the “alignment tax.” In biomedicine, this manifests as a conflict between the need for rigorous, encyclopedic factual accuracy and the requirement for flexible, user-friendly instruction following. Existing methods, primarily relying on Supervised Fine-Tuning (SFT) or standard Reinforcement Learning from Human Feedback (RLHF), often fail to navigate this Pareto frontier, resulting in models that are either knowledgeable but rigid, or chatty but hallucination-prone. In this paper, we propose BalancedBio, a novel alignment framework that explicitly models and optimizes conflicting objectives. While utilizing standard high-quality instruction data, our contribution lies in the algorithmic innovation: (1) We introduce Capability-Aware Group Relative Policy Optimization (GRPO), which eliminates the need for a value network and reduces gradient variance in high-entropy reasoning tasks; (2) We propose a Dynamic Hybrid Reward Mechanism that adaptively balances domain correctness, reasoning validity, and format compliance during training; (3) We provide a theoretical analysis demonstrating how our method enforces gradient orthogonality to mitigate catastrophic forgetting. Extensive experiments on BIOMED-MMLU, MedQA, and IFEval show that BalancedBio-7B achieves state-of-the-art performance, surpassing Med-PaLM-8B in domain tasks while maintaining robust general instruction-following capabilities. We will release our model and partial data.
%U https://aclanthology.org/2026.emnlp-main.2556/
%P 58101-58115

Markdown (Informal)

[BalancedBio: Mitigating the Alignment Tax in Biomedical LLMs via Gradient Orthogonality and GRPO](https://aclanthology.org/2026.emnlp-main.2556/) (Wu et al., EMNLP 2026)

• BalancedBio: Mitigating the Alignment Tax in Biomedical LLMs via Gradient Orthogonality and GRPO (Wu et al., EMNLP 2026)

ACL

• Wentao Wu, Hanmeng Zhong, Linqing Chen, and Peng Zhou. 2026. BalancedBio: Mitigating the Alignment Tax in Biomedical LLMs via Gradient Orthogonality and GRPO. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 58101–58115, Budapest, Hungary. Association for Computational Linguistics.