@inproceedings{veigel-mayer-2026-conformbench,
title = "{C}on{F}orm{B}ench: Evaluating Evidence-Supported Record Maintenance in Conversational Form Filling",
author = "Veigel, Timothy and
Mayer, Sven",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.2137/",
pages = "48575--48599",
ISBN = "979-8-89176-459-0",
abstract = "Conversational form-filling systems must update structured records across dialogue turns. A user may add, correct, or retract information, while the existing record may itself contain stale or unsupported values. We call correct maintenance of this state record commitment fidelity: after each turn, the complete record should contain all and only commitments supported by the schema and dialogue history. We introduce ConFormBench, a diagnostic benchmark that tests this capability. Each item provides a form schema, a prior record, dialogue history, and a new user utterance; the system must return the complete updated record. ConFormBench contains 180 source-grounded synthetic update turns across four institutional form families, covering 339 scoreable fields and 1,683 changed leaf-field commitments. We mechanically validate all items and audit a coverage-controlled 60-item subset through blind independent reconstruction and discrepancy adjudication. We also audit the evaluator against the adjudicated labels. On the 144-turn held-out split, five systems achieve 95.1-97.0{\%} all-field accuracy, yet only 18.1-33.3{\%} of records are entirely correct. Errors concentrate on clearing values, using history, coordinating gates, routing repeated instances, and repairing prior-state errors. Thus, field-level scores substantially overstate record-maintenance reliability."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="veigel-mayer-2026-conformbench">
<titleInfo>
<title>ConFormBench: Evaluating Evidence-Supported Record Maintenance in Conversational Form Filling</title>
</titleInfo>
<name type="personal">
<namePart type="given">Timothy</namePart>
<namePart type="family">Veigel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sven</namePart>
<namePart type="family">Mayer</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Conversational form-filling systems must update structured records across dialogue turns. A user may add, correct, or retract information, while the existing record may itself contain stale or unsupported values. We call correct maintenance of this state record commitment fidelity: after each turn, the complete record should contain all and only commitments supported by the schema and dialogue history. We introduce ConFormBench, a diagnostic benchmark that tests this capability. Each item provides a form schema, a prior record, dialogue history, and a new user utterance; the system must return the complete updated record. ConFormBench contains 180 source-grounded synthetic update turns across four institutional form families, covering 339 scoreable fields and 1,683 changed leaf-field commitments. We mechanically validate all items and audit a coverage-controlled 60-item subset through blind independent reconstruction and discrepancy adjudication. We also audit the evaluator against the adjudicated labels. On the 144-turn held-out split, five systems achieve 95.1-97.0% all-field accuracy, yet only 18.1-33.3% of records are entirely correct. Errors concentrate on clearing values, using history, coordinating gates, routing repeated instances, and repairing prior-state errors. Thus, field-level scores substantially overstate record-maintenance reliability.</abstract>
<identifier type="citekey">veigel-mayer-2026-conformbench</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2137/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>48575</start>
<end>48599</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T ConFormBench: Evaluating Evidence-Supported Record Maintenance in Conversational Form Filling
%A Veigel, Timothy
%A Mayer, Sven
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F veigel-mayer-2026-conformbench
%X Conversational form-filling systems must update structured records across dialogue turns. A user may add, correct, or retract information, while the existing record may itself contain stale or unsupported values. We call correct maintenance of this state record commitment fidelity: after each turn, the complete record should contain all and only commitments supported by the schema and dialogue history. We introduce ConFormBench, a diagnostic benchmark that tests this capability. Each item provides a form schema, a prior record, dialogue history, and a new user utterance; the system must return the complete updated record. ConFormBench contains 180 source-grounded synthetic update turns across four institutional form families, covering 339 scoreable fields and 1,683 changed leaf-field commitments. We mechanically validate all items and audit a coverage-controlled 60-item subset through blind independent reconstruction and discrepancy adjudication. We also audit the evaluator against the adjudicated labels. On the 144-turn held-out split, five systems achieve 95.1-97.0% all-field accuracy, yet only 18.1-33.3% of records are entirely correct. Errors concentrate on clearing values, using history, coordinating gates, routing repeated instances, and repairing prior-state errors. Thus, field-level scores substantially overstate record-maintenance reliability.
%U https://aclanthology.org/2026.emnlp-main.2137/
%P 48575-48599
Markdown (Informal)
[ConFormBench: Evaluating Evidence-Supported Record Maintenance in Conversational Form Filling](https://aclanthology.org/2026.emnlp-main.2137/) (Veigel & Mayer, EMNLP 2026)
ACL