@inproceedings{alijanpour-shalmani-etal-2026-database,
title = "When the Database Fails: Prompting {LLM} Dialogue Agents for Safe Recovery in Task-Oriented Dialogue",
author = "Alijanpour Shalmani, Mohammad and
Rezvani Boroujeni, Alale and
Yuan, Jiann S.",
editor = "Choi, Jinho D. and
Chen, Yun-Nung and
Funakoshi, Kotaro and
Emami, Ali",
booktitle = "Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue",
month = aug,
year = "2026",
address = "Atlanta, Georgia, USA",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.sigdial-1.57/",
pages = "815--820",
abstract = "Large language models used in task-oriented dialogue often produce fluent but unsafe responses when backend database calls fail, return empty results, or surface mismatched information, inventing venues, confirmations, or booking details not grounded in the database. We study a lightweight prompting-based recovery approach that improves robustness without retraining or additional model calls. We compare three response strategies, including a guided recovery prompt conditioned on structured database status, across six open-weight model families (DeepSeek-R1, Gemma-2, Llama-3, Mistral, Phi-3, and Qwen-2.5) and four database conditions: empty result, wrong-domain retrieval, API error, and clean retrieval. Using fault-injected benchmarks built on two structurally different datasets, MultiWOZ 2.2 (5 domains) and SGD (20 domains), we find that naive agents hallucinate on 30.5{\%} of failure turns on MultiWOZ and 20.9{\%} on SGD. Our Guided-Retry strategy reduces hallucination by 50{\%} on MultiWOZ (30.5{\textrightarrow}15.3{\%}) and by 42{\%} on SGD (20.9{\textrightarrow}12.2{\%}) without retraining. However, residual hallucination remains substantial (6{--}37{\%} across models), with wrong-domain failures the hardest case. Results are consistent across both datasets and all six model families, and human annotation shows substantial agreement while supporting the validity of the automatic commitment-safety metric."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="alijanpour-shalmani-etal-2026-database">
<titleInfo>
<title>When the Database Fails: Prompting LLM Dialogue Agents for Safe Recovery in Task-Oriented Dialogue</title>
</titleInfo>
<name type="personal">
<namePart type="given">Mohammad</namePart>
<namePart type="family">Alijanpour Shalmani</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alale</namePart>
<namePart type="family">Rezvani Boroujeni</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiann</namePart>
<namePart type="given">S</namePart>
<namePart type="family">Yuan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-08</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jinho</namePart>
<namePart type="given">D</namePart>
<namePart type="family">Choi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yun-Nung</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kotaro</namePart>
<namePart type="family">Funakoshi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ali</namePart>
<namePart type="family">Emami</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Atlanta, Georgia, USA</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Large language models used in task-oriented dialogue often produce fluent but unsafe responses when backend database calls fail, return empty results, or surface mismatched information, inventing venues, confirmations, or booking details not grounded in the database. We study a lightweight prompting-based recovery approach that improves robustness without retraining or additional model calls. We compare three response strategies, including a guided recovery prompt conditioned on structured database status, across six open-weight model families (DeepSeek-R1, Gemma-2, Llama-3, Mistral, Phi-3, and Qwen-2.5) and four database conditions: empty result, wrong-domain retrieval, API error, and clean retrieval. Using fault-injected benchmarks built on two structurally different datasets, MultiWOZ 2.2 (5 domains) and SGD (20 domains), we find that naive agents hallucinate on 30.5% of failure turns on MultiWOZ and 20.9% on SGD. Our Guided-Retry strategy reduces hallucination by 50% on MultiWOZ (30.5→15.3%) and by 42% on SGD (20.9→12.2%) without retraining. However, residual hallucination remains substantial (6–37% across models), with wrong-domain failures the hardest case. Results are consistent across both datasets and all six model families, and human annotation shows substantial agreement while supporting the validity of the automatic commitment-safety metric.</abstract>
<identifier type="citekey">alijanpour-shalmani-etal-2026-database</identifier>
<location>
<url>https://aclanthology.org/2026.sigdial-1.57/</url>
</location>
<part>
<date>2026-08</date>
<extent unit="page">
<start>815</start>
<end>820</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T When the Database Fails: Prompting LLM Dialogue Agents for Safe Recovery in Task-Oriented Dialogue
%A Alijanpour Shalmani, Mohammad
%A Rezvani Boroujeni, Alale
%A Yuan, Jiann S.
%Y Choi, Jinho D.
%Y Chen, Yun-Nung
%Y Funakoshi, Kotaro
%Y Emami, Ali
%S Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue
%D 2026
%8 August
%I Association for Computational Linguistics
%C Atlanta, Georgia, USA
%F alijanpour-shalmani-etal-2026-database
%X Large language models used in task-oriented dialogue often produce fluent but unsafe responses when backend database calls fail, return empty results, or surface mismatched information, inventing venues, confirmations, or booking details not grounded in the database. We study a lightweight prompting-based recovery approach that improves robustness without retraining or additional model calls. We compare three response strategies, including a guided recovery prompt conditioned on structured database status, across six open-weight model families (DeepSeek-R1, Gemma-2, Llama-3, Mistral, Phi-3, and Qwen-2.5) and four database conditions: empty result, wrong-domain retrieval, API error, and clean retrieval. Using fault-injected benchmarks built on two structurally different datasets, MultiWOZ 2.2 (5 domains) and SGD (20 domains), we find that naive agents hallucinate on 30.5% of failure turns on MultiWOZ and 20.9% on SGD. Our Guided-Retry strategy reduces hallucination by 50% on MultiWOZ (30.5→15.3%) and by 42% on SGD (20.9→12.2%) without retraining. However, residual hallucination remains substantial (6–37% across models), with wrong-domain failures the hardest case. Results are consistent across both datasets and all six model families, and human annotation shows substantial agreement while supporting the validity of the automatic commitment-safety metric.
%U https://aclanthology.org/2026.sigdial-1.57/
%P 815-820
Markdown (Informal)
[When the Database Fails: Prompting LLM Dialogue Agents for Safe Recovery in Task-Oriented Dialogue](https://aclanthology.org/2026.sigdial-1.57/) (Alijanpour Shalmani et al., SIGDIAL 2026)
ACL