<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="lai-si-2026-llms">
    <titleInfo>
        <title>Do LLMs Respect Concept Boundaries? A Bilingual Diagnostic Benchmark</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Renjun</namePart>
        <namePart type="family">Lai</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yixuan</namePart>
        <namePart type="family">Si</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>Large language models are increasingly used as natural-language interfaces for structured information processing and agentic workflows, where user-specified constraints can determine downstream extraction, selection, tool use, and action planning. This paper studies whether models can execute natural-language concept-set constraints over candidate objects: given inclusion and exclusion conditions, the model must decide which candidates fall inside the intended concept boundary. We introduce a dual-resource diagnostic benchmark built from WordNet and BabelNet. The benchmark fixes backend constraints, candidate realizations, and gold answers before prompt rendering, and evaluates three task views: single-object judgment, exact candidate-set construction, and satisfiability judgment. The main evaluation set contains 202,500 rendered requests. Across 11 scored model runs, overall accuracy ranges from 43.70% to 56.36%, showing that the benchmark remains difficult for current strong models. Exact candidate-set construction yields consistently lower accuracy under stricter set-construction requirements, and recall diagnostics reveal asymmetric effects of constraint direction: increasing exclusion constraints lowers positive recall and raises negative recall, while increasing inclusion branches mainly erodes false-case rejection. These results suggest that concept-boundary execution is a measurable weakness under this diagnostic setup and that evaluating it requires factor-level diagnostics beyond aggregate accuracy.</abstract>
    <identifier type="citekey">lai-si-2026-llms</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-aacl.89/</url>
    </location>
    <part>
        <date>2026-11</date>
        <extent unit="page">
            <start>1797</start>
            <end>1814</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Do LLMs Respect Concept Boundaries? A Bilingual Diagnostic Benchmark
%A Lai, Renjun
%A Si, Yixuan
%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 lai-si-2026-llms
%X Large language models are increasingly used as natural-language interfaces for structured information processing and agentic workflows, where user-specified constraints can determine downstream extraction, selection, tool use, and action planning. This paper studies whether models can execute natural-language concept-set constraints over candidate objects: given inclusion and exclusion conditions, the model must decide which candidates fall inside the intended concept boundary. We introduce a dual-resource diagnostic benchmark built from WordNet and BabelNet. The benchmark fixes backend constraints, candidate realizations, and gold answers before prompt rendering, and evaluates three task views: single-object judgment, exact candidate-set construction, and satisfiability judgment. The main evaluation set contains 202,500 rendered requests. Across 11 scored model runs, overall accuracy ranges from 43.70% to 56.36%, showing that the benchmark remains difficult for current strong models. Exact candidate-set construction yields consistently lower accuracy under stricter set-construction requirements, and recall diagnostics reveal asymmetric effects of constraint direction: increasing exclusion constraints lowers positive recall and raises negative recall, while increasing inclusion branches mainly erodes false-case rejection. These results suggest that concept-boundary execution is a measurable weakness under this diagnostic setup and that evaluating it requires factor-level diagnostics beyond aggregate accuracy.
%U https://aclanthology.org/2026.findings-aacl.89/
%P 1797-1814

Markdown (Informal)

[Do LLMs Respect Concept Boundaries? A Bilingual Diagnostic Benchmark](https://aclanthology.org/2026.findings-aacl.89/) (Lai & Si, Findings 2026)

• Do LLMs Respect Concept Boundaries? A Bilingual Diagnostic Benchmark (Lai & Si, Findings 2026)

ACL

• Renjun Lai and Yixuan Si. 2026. Do LLMs Respect Concept Boundaries? A Bilingual Diagnostic Benchmark. In Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026, pages 1797–1814, Hengqin, China. Association for Computational Linguistics.