@inproceedings{wang-etal-2026-diagnosing-repairing,
title = "Diagnosing and Repairing Local Reasoning Preference Collapse in Medical Chain-of-Thought Distillation",
author = "Wang, RuiQing and
Tan, Xiaoyu and
Wang, Xiyao and
Chen, Jue and
Qiu, Xihe",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.964/",
doi = "10.18653/2026.findings-emnlp.964",
pages = "20390--20412",
ISBN = "979-8-89176-462-0",
abstract = "Chain-of-thought (CoT) distillation is a common strategy for transferring reasoning abilities from large teacher models to smaller student models, but in medical question answering, a student may produce plausible rationales while locally drifting toward clinically incorrect reasoning paths. This makes final-answer supervision and uniform rationale imitation insufficient for identifying where reasoning fails. We study this problem as local reasoning preference collapse in structured medical CoT distillation and propose Answer-Calibrated Collapse-Aware Preference Distillation (ACPD). ACPD first exposes the student{'}s local continuation preferences through teacher-guided verbalized candidate construction and measures domain-level collapse with a Local Collapse Margin (LCM). It then selectively repairs collapse-prone domains by contrasting the student{'}s most preferred teacher-approved and teacher-disapproved continuations. To ensure that local repair improves decision making rather than only rationale form, ACPD further introduces answer calibration, which anchors the full structured reasoning prefix to the correct final choice. Experiments on medical multiple-choice reasoning benchmarks show that ACPD improves both answer accuracy and collapse-sensitive repair metrics over strong distillation baselines while maintaining competitive out-of-distribution generalization. These results suggest that effective medical CoT distillation should not only transfer teacher rationales, but also diagnose and repair where the student{'}s reasoning preferences locally diverge from medically reliable reasoning paths."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wang-etal-2026-diagnosing-repairing">
<titleInfo>
<title>Diagnosing and Repairing Local Reasoning Preference Collapse in Medical Chain-of-Thought Distillation</title>
</titleInfo>
<name type="personal">
<namePart type="given">RuiQing</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaoyu</namePart>
<namePart type="family">Tan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiyao</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jue</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xihe</namePart>
<namePart type="family">Qiu</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>Findings of the Association for Computational Linguistics: EMNLP 2026</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">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Chain-of-thought (CoT) distillation is a common strategy for transferring reasoning abilities from large teacher models to smaller student models, but in medical question answering, a student may produce plausible rationales while locally drifting toward clinically incorrect reasoning paths. This makes final-answer supervision and uniform rationale imitation insufficient for identifying where reasoning fails. We study this problem as local reasoning preference collapse in structured medical CoT distillation and propose Answer-Calibrated Collapse-Aware Preference Distillation (ACPD). ACPD first exposes the student’s local continuation preferences through teacher-guided verbalized candidate construction and measures domain-level collapse with a Local Collapse Margin (LCM). It then selectively repairs collapse-prone domains by contrasting the student’s most preferred teacher-approved and teacher-disapproved continuations. To ensure that local repair improves decision making rather than only rationale form, ACPD further introduces answer calibration, which anchors the full structured reasoning prefix to the correct final choice. Experiments on medical multiple-choice reasoning benchmarks show that ACPD improves both answer accuracy and collapse-sensitive repair metrics over strong distillation baselines while maintaining competitive out-of-distribution generalization. These results suggest that effective medical CoT distillation should not only transfer teacher rationales, but also diagnose and repair where the student’s reasoning preferences locally diverge from medically reliable reasoning paths.</abstract>
<identifier type="citekey">wang-etal-2026-diagnosing-repairing</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.964</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.964/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>20390</start>
<end>20412</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Diagnosing and Repairing Local Reasoning Preference Collapse in Medical Chain-of-Thought Distillation
%A Wang, RuiQing
%A Tan, Xiaoyu
%A Wang, Xiyao
%A Chen, Jue
%A Qiu, Xihe
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F wang-etal-2026-diagnosing-repairing
%X Chain-of-thought (CoT) distillation is a common strategy for transferring reasoning abilities from large teacher models to smaller student models, but in medical question answering, a student may produce plausible rationales while locally drifting toward clinically incorrect reasoning paths. This makes final-answer supervision and uniform rationale imitation insufficient for identifying where reasoning fails. We study this problem as local reasoning preference collapse in structured medical CoT distillation and propose Answer-Calibrated Collapse-Aware Preference Distillation (ACPD). ACPD first exposes the student’s local continuation preferences through teacher-guided verbalized candidate construction and measures domain-level collapse with a Local Collapse Margin (LCM). It then selectively repairs collapse-prone domains by contrasting the student’s most preferred teacher-approved and teacher-disapproved continuations. To ensure that local repair improves decision making rather than only rationale form, ACPD further introduces answer calibration, which anchors the full structured reasoning prefix to the correct final choice. Experiments on medical multiple-choice reasoning benchmarks show that ACPD improves both answer accuracy and collapse-sensitive repair metrics over strong distillation baselines while maintaining competitive out-of-distribution generalization. These results suggest that effective medical CoT distillation should not only transfer teacher rationales, but also diagnose and repair where the student’s reasoning preferences locally diverge from medically reliable reasoning paths.
%R 10.18653/2026.findings-emnlp.964
%U https://aclanthology.org/2026.findings-emnlp.964/
%U https://doi.org/10.18653/2026.findings-emnlp.964
%P 20390-20412
Markdown (Informal)
[Diagnosing and Repairing Local Reasoning Preference Collapse in Medical Chain-of-Thought Distillation](https://aclanthology.org/2026.findings-emnlp.964/) (Wang et al., Findings 2026)
ACL