@inproceedings{liang-etal-2026-harm,
title = "From Harm to Help: Turning Reasoning In-Context Demos into Assets for Reasoning {LM}s",
author = "Liang, Weida and
Fu, Zihang and
Wang, Haonan and
Nie, Zheng and
Zhang, Yifan and
Tong, Yao and
Zhu, Tongyao and
Jiang, Hao and
Li, Chuang and
Wu, Jiaying and
Kawaguchi, Kenji",
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.66/",
pages = "1333--1355",
ISBN = "979-8-89176-462-0",
abstract = "Recent reasoning LLMs (RLMs), especially those trained with verifier-based reinforcement learning, often perform worse with few-shot CoT than with direct answering. We revisit this paradox using high-quality reasoning traces from DeepSeek-R1 as demonstrations and find that adding more exemplars consistently degrades accuracy, even when demonstrations are optimal. A detailed analysis reveals two mechanisms behind this decline: (i) *semantic misguidance*, where high textual similarity leads the model to treat the target as the same as the exemplar and to copy intermediate steps verbatim; and (ii) *strategy transfer failure*, where the model struggles to extract useful reasoning strategies and apply them to target questions. Guided by these findings, we introduce Insight-to-Solve (I2S), which converts demonstrations into target-conditioned insights before solving, and I2S+, which adds iterative refinement. Across open- and closed-source models, both variants outperform direct answering and sequential test-time baselines. On a frozen 500-problem Omni-MATH transfer set, I2S+ improves R1-Distill-Qwen-7B and Qwen3-14B by $+4.2$ and $+4.0$ points, respectively. The same prompt stack also transfers to GPT-4.1, o1-mini, and R1-Distill-Llama-8B, showing that reasoning demonstrations can be converted from liabilities into useful test-time guidance."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liang-etal-2026-harm">
<titleInfo>
<title>From Harm to Help: Turning Reasoning In-Context Demos into Assets for Reasoning LMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Weida</namePart>
<namePart type="family">Liang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zihang</namePart>
<namePart type="family">Fu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haonan</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zheng</namePart>
<namePart type="family">Nie</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yifan</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yao</namePart>
<namePart type="family">Tong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tongyao</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hao</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chuang</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiaying</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kenji</namePart>
<namePart type="family">Kawaguchi</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>Recent reasoning LLMs (RLMs), especially those trained with verifier-based reinforcement learning, often perform worse with few-shot CoT than with direct answering. We revisit this paradox using high-quality reasoning traces from DeepSeek-R1 as demonstrations and find that adding more exemplars consistently degrades accuracy, even when demonstrations are optimal. A detailed analysis reveals two mechanisms behind this decline: (i) *semantic misguidance*, where high textual similarity leads the model to treat the target as the same as the exemplar and to copy intermediate steps verbatim; and (ii) *strategy transfer failure*, where the model struggles to extract useful reasoning strategies and apply them to target questions. Guided by these findings, we introduce Insight-to-Solve (I2S), which converts demonstrations into target-conditioned insights before solving, and I2S+, which adds iterative refinement. Across open- and closed-source models, both variants outperform direct answering and sequential test-time baselines. On a frozen 500-problem Omni-MATH transfer set, I2S+ improves R1-Distill-Qwen-7B and Qwen3-14B by +4.2 and +4.0 points, respectively. The same prompt stack also transfers to GPT-4.1, o1-mini, and R1-Distill-Llama-8B, showing that reasoning demonstrations can be converted from liabilities into useful test-time guidance.</abstract>
<identifier type="citekey">liang-etal-2026-harm</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.66/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>1333</start>
<end>1355</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T From Harm to Help: Turning Reasoning In-Context Demos into Assets for Reasoning LMs
%A Liang, Weida
%A Fu, Zihang
%A Wang, Haonan
%A Nie, Zheng
%A Zhang, Yifan
%A Tong, Yao
%A Zhu, Tongyao
%A Jiang, Hao
%A Li, Chuang
%A Wu, Jiaying
%A Kawaguchi, Kenji
%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 liang-etal-2026-harm
%X Recent reasoning LLMs (RLMs), especially those trained with verifier-based reinforcement learning, often perform worse with few-shot CoT than with direct answering. We revisit this paradox using high-quality reasoning traces from DeepSeek-R1 as demonstrations and find that adding more exemplars consistently degrades accuracy, even when demonstrations are optimal. A detailed analysis reveals two mechanisms behind this decline: (i) *semantic misguidance*, where high textual similarity leads the model to treat the target as the same as the exemplar and to copy intermediate steps verbatim; and (ii) *strategy transfer failure*, where the model struggles to extract useful reasoning strategies and apply them to target questions. Guided by these findings, we introduce Insight-to-Solve (I2S), which converts demonstrations into target-conditioned insights before solving, and I2S+, which adds iterative refinement. Across open- and closed-source models, both variants outperform direct answering and sequential test-time baselines. On a frozen 500-problem Omni-MATH transfer set, I2S+ improves R1-Distill-Qwen-7B and Qwen3-14B by +4.2 and +4.0 points, respectively. The same prompt stack also transfers to GPT-4.1, o1-mini, and R1-Distill-Llama-8B, showing that reasoning demonstrations can be converted from liabilities into useful test-time guidance.
%U https://aclanthology.org/2026.findings-emnlp.66/
%P 1333-1355
Markdown (Informal)
[From Harm to Help: Turning Reasoning In-Context Demos into Assets for Reasoning LMs](https://aclanthology.org/2026.findings-emnlp.66/) (Liang et al., Findings 2026)
ACL
- Weida Liang, Zihang Fu, Haonan Wang, Zheng Nie, Yifan Zhang, Yao Tong, Tongyao Zhu, Hao Jiang, Chuang Li, Jiaying Wu, and Kenji Kawaguchi. 2026. From Harm to Help: Turning Reasoning In-Context Demos into Assets for Reasoning LMs. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 1333–1355, Budapest, Hungary. Association for Computational Linguistics.