@inproceedings{nursal-etal-2026-supervision,
title = "When more supervision hurts: reward design for structured biomedical reasoning in 4{B} models",
author = "Nursal, Batuhan and
Ye, Christophe and
Guo, Xinyu and
He, Xin and
Mitchell, Cassie S.",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.145/",
pages = "2851--2866",
ISBN = "979-8-89176-444-6",
abstract = "Reinforcement learning has become a common approach for improving language-model reasoning, but its effectiveness depends heavily on reward design, especially for compact domain-specialized models. We study structured biomedical question answering with two 4B-scale models, MedGemma-1.5-4B-it and Qwen3-4B-Thinking-2507, using a two-stage SFT+GRPO pipeline that trains models to produce tagged think/answer outputs. We compare three reward schemes: a simple format + accuracy reward, a depth incentivized reward that encourages multiple reasoning cycles, and a gated LLM-judge reward for intermediate reasoning milestones. Across five biomedical QA benchmarks, the simplest reward is the most robust: it substantially improves MedGemma in the structured-output setting, while depth and LLM-judge rewards often reduce accuracy. For Qwen3, SFT or native long-budget inference can outperform GRPO on several benchmarks, showing that reasoning-tuned models may not benefit from additional structural reinforcement. Our results suggest that, for lightweight biomedical QA, reward simplicity and model priors are more important than increasingly elaborate intermediate reasoning supervision. The codebase for this study is publicly available on GitHub: https://github.com/batuinmetz/When-more-supervision-hurts."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="nursal-etal-2026-supervision">
<titleInfo>
<title>When more supervision hurts: reward design for structured biomedical reasoning in 4B models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Batuhan</namePart>
<namePart type="family">Nursal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christophe</namePart>
<namePart type="family">Ye</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinyu</namePart>
<namePart type="family">Guo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xin</namePart>
<namePart type="family">He</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Cassie</namePart>
<namePart type="given">S</namePart>
<namePart type="family">Mitchell</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>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</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">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>Reinforcement learning has become a common approach for improving language-model reasoning, but its effectiveness depends heavily on reward design, especially for compact domain-specialized models. We study structured biomedical question answering with two 4B-scale models, MedGemma-1.5-4B-it and Qwen3-4B-Thinking-2507, using a two-stage SFT+GRPO pipeline that trains models to produce tagged think/answer outputs. We compare three reward schemes: a simple format + accuracy reward, a depth incentivized reward that encourages multiple reasoning cycles, and a gated LLM-judge reward for intermediate reasoning milestones. Across five biomedical QA benchmarks, the simplest reward is the most robust: it substantially improves MedGemma in the structured-output setting, while depth and LLM-judge rewards often reduce accuracy. For Qwen3, SFT or native long-budget inference can outperform GRPO on several benchmarks, showing that reasoning-tuned models may not benefit from additional structural reinforcement. Our results suggest that, for lightweight biomedical QA, reward simplicity and model priors are more important than increasingly elaborate intermediate reasoning supervision. The codebase for this study is publicly available on GitHub: https://github.com/batuinmetz/When-more-supervision-hurts.</abstract>
<identifier type="citekey">nursal-etal-2026-supervision</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.145/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>2851</start>
<end>2866</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T When more supervision hurts: reward design for structured biomedical reasoning in 4B models
%A Nursal, Batuhan
%A Ye, Christophe
%A Guo, Xinyu
%A He, Xin
%A Mitchell, Cassie S.
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F nursal-etal-2026-supervision
%X Reinforcement learning has become a common approach for improving language-model reasoning, but its effectiveness depends heavily on reward design, especially for compact domain-specialized models. We study structured biomedical question answering with two 4B-scale models, MedGemma-1.5-4B-it and Qwen3-4B-Thinking-2507, using a two-stage SFT+GRPO pipeline that trains models to produce tagged think/answer outputs. We compare three reward schemes: a simple format + accuracy reward, a depth incentivized reward that encourages multiple reasoning cycles, and a gated LLM-judge reward for intermediate reasoning milestones. Across five biomedical QA benchmarks, the simplest reward is the most robust: it substantially improves MedGemma in the structured-output setting, while depth and LLM-judge rewards often reduce accuracy. For Qwen3, SFT or native long-budget inference can outperform GRPO on several benchmarks, showing that reasoning-tuned models may not benefit from additional structural reinforcement. Our results suggest that, for lightweight biomedical QA, reward simplicity and model priors are more important than increasingly elaborate intermediate reasoning supervision. The codebase for this study is publicly available on GitHub: https://github.com/batuinmetz/When-more-supervision-hurts.
%U https://aclanthology.org/2026.findings-aacl.145/
%P 2851-2866
Markdown (Informal)
[When more supervision hurts: reward design for structured biomedical reasoning in 4B models](https://aclanthology.org/2026.findings-aacl.145/) (Nursal et al., Findings 2026)
ACL