@inproceedings{zhao-etal-2026-calibration,
title = "Calibration is the Bottleneck: An Action-Class Diagnostic of Multi-Turn Tool-Calling",
author = "Zhao, Kangjia and
Li, Jiajun and
Shen, Haozhan and
Chow, Wei and
Li, Linfeng and
Song, Hang and
Kong, Lingdong and
Zhi, Chen and
Zhao, Tiancheng and
Liu, Songhua and
Yin, Jianwei",
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.305/",
pages = "6463--6480",
ISBN = "979-8-89176-462-0",
abstract = "Multi-turn tool calling is a core evaluation scenario for large language model (LLM) agents. On public tool-calling benchmarks, open-weight models now approach or even surpass closed-source frontier models in aggregate accuracy. However, aggregate accuracy averages over many different multi-turn situations and obscures whether progress is balanced across them, masking failure redistribution. We propose an action-class-oriented diagnostic framework that decomposes failures into two orthogonal dimensions: capability and calibration. The framework operates over a four-class action space (tool{\_}call/ask/refuse/confirm) and introduces a self-revealing upper bound, Acc {\ensuremath{\leq}} GAR (Gold Action Recall), whose two violation directions separately localize action-class miscalibration and sub-action failures. We validate it on an extensive panel of tool-calling models that includes both open-weight families and closed-source frontier systems, across multiple multi-turn tool-calling benchmarks for cross-benchmark robustness. Across our panel the bottleneck is calibration, not capability: every model can produce all four classes, but which one it fires in which situation is highly model-specific. On no-tool-call scenarios, a top-ranked family{'}s GAR is roughly one-quarter that of size-matched peers, a 64.5 percentage point (pp) gap, indicating its ranking reflects benchmark-distribution fitting rather than genuine in-context calibration. We further show the calibration profile can be reshaped through context manipulation alone, with the reshaping direction strongly model-specific. We recommend that multi-turn tool-calling evaluations report per-category (GAR, tool-call-emit rate, Acc) triples, exposing the calibration structure aggregate accuracy hides. Our framework provides a falsifiable, reproducible diagnostic foundation for evaluation protocol design, training objective alignment, and inference-time calibration research."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhao-etal-2026-calibration">
<titleInfo>
<title>Calibration is the Bottleneck: An Action-Class Diagnostic of Multi-Turn Tool-Calling</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kangjia</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiajun</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haozhan</namePart>
<namePart type="family">Shen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Chow</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Linfeng</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hang</namePart>
<namePart type="family">Song</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lingdong</namePart>
<namePart type="family">Kong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chen</namePart>
<namePart type="family">Zhi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tiancheng</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Songhua</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jianwei</namePart>
<namePart type="family">Yin</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>Multi-turn tool calling is a core evaluation scenario for large language model (LLM) agents. On public tool-calling benchmarks, open-weight models now approach or even surpass closed-source frontier models in aggregate accuracy. However, aggregate accuracy averages over many different multi-turn situations and obscures whether progress is balanced across them, masking failure redistribution. We propose an action-class-oriented diagnostic framework that decomposes failures into two orthogonal dimensions: capability and calibration. The framework operates over a four-class action space (tool_call/ask/refuse/confirm) and introduces a self-revealing upper bound, Acc \ensuremathłeq GAR (Gold Action Recall), whose two violation directions separately localize action-class miscalibration and sub-action failures. We validate it on an extensive panel of tool-calling models that includes both open-weight families and closed-source frontier systems, across multiple multi-turn tool-calling benchmarks for cross-benchmark robustness. Across our panel the bottleneck is calibration, not capability: every model can produce all four classes, but which one it fires in which situation is highly model-specific. On no-tool-call scenarios, a top-ranked family’s GAR is roughly one-quarter that of size-matched peers, a 64.5 percentage point (pp) gap, indicating its ranking reflects benchmark-distribution fitting rather than genuine in-context calibration. We further show the calibration profile can be reshaped through context manipulation alone, with the reshaping direction strongly model-specific. We recommend that multi-turn tool-calling evaluations report per-category (GAR, tool-call-emit rate, Acc) triples, exposing the calibration structure aggregate accuracy hides. Our framework provides a falsifiable, reproducible diagnostic foundation for evaluation protocol design, training objective alignment, and inference-time calibration research.</abstract>
<identifier type="citekey">zhao-etal-2026-calibration</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.305/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>6463</start>
<end>6480</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Calibration is the Bottleneck: An Action-Class Diagnostic of Multi-Turn Tool-Calling
%A Zhao, Kangjia
%A Li, Jiajun
%A Shen, Haozhan
%A Chow, Wei
%A Li, Linfeng
%A Song, Hang
%A Kong, Lingdong
%A Zhi, Chen
%A Zhao, Tiancheng
%A Liu, Songhua
%A Yin, Jianwei
%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 zhao-etal-2026-calibration
%X Multi-turn tool calling is a core evaluation scenario for large language model (LLM) agents. On public tool-calling benchmarks, open-weight models now approach or even surpass closed-source frontier models in aggregate accuracy. However, aggregate accuracy averages over many different multi-turn situations and obscures whether progress is balanced across them, masking failure redistribution. We propose an action-class-oriented diagnostic framework that decomposes failures into two orthogonal dimensions: capability and calibration. The framework operates over a four-class action space (tool_call/ask/refuse/confirm) and introduces a self-revealing upper bound, Acc \ensuremathłeq GAR (Gold Action Recall), whose two violation directions separately localize action-class miscalibration and sub-action failures. We validate it on an extensive panel of tool-calling models that includes both open-weight families and closed-source frontier systems, across multiple multi-turn tool-calling benchmarks for cross-benchmark robustness. Across our panel the bottleneck is calibration, not capability: every model can produce all four classes, but which one it fires in which situation is highly model-specific. On no-tool-call scenarios, a top-ranked family’s GAR is roughly one-quarter that of size-matched peers, a 64.5 percentage point (pp) gap, indicating its ranking reflects benchmark-distribution fitting rather than genuine in-context calibration. We further show the calibration profile can be reshaped through context manipulation alone, with the reshaping direction strongly model-specific. We recommend that multi-turn tool-calling evaluations report per-category (GAR, tool-call-emit rate, Acc) triples, exposing the calibration structure aggregate accuracy hides. Our framework provides a falsifiable, reproducible diagnostic foundation for evaluation protocol design, training objective alignment, and inference-time calibration research.
%U https://aclanthology.org/2026.findings-emnlp.305/
%P 6463-6480
Markdown (Informal)
[Calibration is the Bottleneck: An Action-Class Diagnostic of Multi-Turn Tool-Calling](https://aclanthology.org/2026.findings-emnlp.305/) (Zhao et al., Findings 2026)
ACL
- Kangjia Zhao, Jiajun Li, Haozhan Shen, Wei Chow, Linfeng Li, Hang Song, Lingdong Kong, Chen Zhi, Tiancheng Zhao, Songhua Liu, and Jianwei Yin. 2026. Calibration is the Bottleneck: An Action-Class Diagnostic of Multi-Turn Tool-Calling. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 6463–6480, Budapest, Hungary. Association for Computational Linguistics.