@inproceedings{li-etal-2026-diagnosing-decision,
title = "Diagnosing Decision-Boundary Induction in Few-Shot In-Context Learning",
author = "Li, Mingyang and
Yue, Xinlang and
Li, Muyang and
Liu, Yongqi and
Zhan, Kaiqiao",
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.1150/",
pages = "24345--24358",
ISBN = "979-8-89176-462-0",
abstract = "Few-shot in-context learning is usually evaluated by final accuracy, but correct predictions do not guarantee recovery of the intended task boundary. We formulate few-shot ICL as decision-boundary induction and decompose performance into boundary induction and rule execution. Using controlled binary verification tasks with decomposable boundaries, we diagnose which necessary components are behaviorally enforced and, under chain-of-thought, explicitly recognized. Across three tasks and multiple model families, we show that few-shot ICL often induces component-biased verifiers rather than the full target boundary. Stronger models are more robust to this bias, while chain-of-thought reshapes rather than uniformly improves boundary recovery. An explicit-rule intervention further shows that improvements come from both recovering missing boundary components and executing recognized ones more reliably. These results suggest that evaluating few-shot ICL should go beyond aggregate accuracy and directly test boundary recovery. Code is available at \url{https://github.com/limy805/boundary-icl}."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-etal-2026-diagnosing-decision">
<titleInfo>
<title>Diagnosing Decision-Boundary Induction in Few-Shot In-Context Learning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Mingyang</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinlang</namePart>
<namePart type="family">Yue</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Muyang</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yongqi</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kaiqiao</namePart>
<namePart type="family">Zhan</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>Few-shot in-context learning is usually evaluated by final accuracy, but correct predictions do not guarantee recovery of the intended task boundary. We formulate few-shot ICL as decision-boundary induction and decompose performance into boundary induction and rule execution. Using controlled binary verification tasks with decomposable boundaries, we diagnose which necessary components are behaviorally enforced and, under chain-of-thought, explicitly recognized. Across three tasks and multiple model families, we show that few-shot ICL often induces component-biased verifiers rather than the full target boundary. Stronger models are more robust to this bias, while chain-of-thought reshapes rather than uniformly improves boundary recovery. An explicit-rule intervention further shows that improvements come from both recovering missing boundary components and executing recognized ones more reliably. These results suggest that evaluating few-shot ICL should go beyond aggregate accuracy and directly test boundary recovery. Code is available at https://github.com/limy805/boundary-icl.</abstract>
<identifier type="citekey">li-etal-2026-diagnosing-decision</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1150/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>24345</start>
<end>24358</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Diagnosing Decision-Boundary Induction in Few-Shot In-Context Learning
%A Li, Mingyang
%A Yue, Xinlang
%A Li, Muyang
%A Liu, Yongqi
%A Zhan, Kaiqiao
%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 li-etal-2026-diagnosing-decision
%X Few-shot in-context learning is usually evaluated by final accuracy, but correct predictions do not guarantee recovery of the intended task boundary. We formulate few-shot ICL as decision-boundary induction and decompose performance into boundary induction and rule execution. Using controlled binary verification tasks with decomposable boundaries, we diagnose which necessary components are behaviorally enforced and, under chain-of-thought, explicitly recognized. Across three tasks and multiple model families, we show that few-shot ICL often induces component-biased verifiers rather than the full target boundary. Stronger models are more robust to this bias, while chain-of-thought reshapes rather than uniformly improves boundary recovery. An explicit-rule intervention further shows that improvements come from both recovering missing boundary components and executing recognized ones more reliably. These results suggest that evaluating few-shot ICL should go beyond aggregate accuracy and directly test boundary recovery. Code is available at https://github.com/limy805/boundary-icl.
%U https://aclanthology.org/2026.findings-emnlp.1150/
%P 24345-24358
Markdown (Informal)
[Diagnosing Decision-Boundary Induction in Few-Shot In-Context Learning](https://aclanthology.org/2026.findings-emnlp.1150/) (Li et al., Findings 2026)
ACL