The Art of Repair in Human-Agent Conversations - A Taxonomy of Repair Strategies by Users and LLM-Based Conversational Agents

Sep 11, 2026

Uploaded by Angel

Large Language Models (LLMs) such as ChatGPT increasingly support everyday work, but their outputs can be unreliable, ambiguous, or misleading. This paper examines how users identify and repair troubles so LLM conversational agents “make things right” in situated interactions. Using an ethnomethodological lens, it analyzes 21 real-world chat transcripts across diverse work contexts and derives a taxonomy covering repair initiators, repair stages, and repair processes.

LanguageEnglish

CategoryPresentations

FormatPDF

Pages51

Reading Time18 min

AccessFree to download

Abstract

Introduction

LLM-based conversational agents and interaction risks

Repair troubles and related research background

University of Siegen University of Siegen Hochschule Bonn-Rhein-Sieg University of Applied Sciences University of Siegen University of Siegen University of Siegen September 2nd, 2025 [https://doi.org/10.21203/rs.3.rs-7301450/v1](https://doi.org/10.21203/rs.3.rs-7301450/v1) 􀁱 􀅏 This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License No competing interests reported. The Art of Repair in Human-Agent Conversations: A Taxonomy of Repair Strategies by Users and LLM-Based Conversational Agents Gunnar Gunnar 1, 2 , Delong Korus-Du 1, 2*, Alexander Boden 2 , Peter Tolmie 1 , Dave Randall 1 , Md Shajalal 1, 2 1 Department of Human-Computer Interaction, University of Siegen, Siegen,Germany. 2Verbraucherinformatik Research Group, Bonn-Rhein-Sieg University of Applied Science, Sankt Augustin, Germany. *Corresponding author(s) . E-mail(s): delong.du@uni-siegen.de; Contributing authors: [gunnar.stevens@uni-siegen.de](gunnar.stevens@uni-siegen.de) ; alexander.boden@h-brs.de; [peter.tolmie@gmail.com](peter.tolmie@gmail.com) ; [daverandall2008@gmail.com](daverandall2008@gmail.com) ; [md.shajalal@uni-siegen.de](md.shajalal@uni-siegen.de) ; Abstract Large Language Models (LLMs) like ChatGPT are increasingly embedded in everyday tasks of various professions, yet LLMs’ outputs often remain unreliable, ambiguous, or misleading. This paper exploreshow users identify and repair troubles to make LLMs make things right in situated interactions of various contexts. Using an ethnomethodological lens, we examine 21 real-world chat transcripts across diverse work contexts. Our analysis reveals a broad repertoire of repair practices, including factual corrections, stylistic re􀀌nements, implicit signals, and strategic reframings. The 􀀌ndings challenge the view that users’ repair work on LLM outputs is merely a response to system failure. Instead, our 􀀌ndings present a taxonomy of repair work of both users and conversational agents, comprising 6 types of repair initiators (errors, dissatisfactions, apologies, shortcomings, implicit signals, and contextualization), 3 stages of repair elements (6 types of trouble classi􀀌cation, 3 types of trouble speci􀀌cation, and 7 types of trouble management), and 3 types of repair processes (incremental, grounding, and validating) . These repair categories demonstrate the core of human-agent collaboration: meaning and correctness are not pre-given but are achieved through situated work for all practical purposes. By treating trouble as an ordinary part of collaborative work, we highlight the need to design for user repair interaction alongside improving model reliability. These 􀀌ndings contribute to ongoing debates in HCI and CSCW around the accountability, intelligibility, and co-construction of meaning in human-AI interaction regarding LLMs applications. 1 Introduction Large Language Model (LLM)-based conversational agents, such as ChatGPT marking a significant step in the evolution of dialogue-like interaction (Moore and Arar, 2019) . Unlike previous rule-based or intentdriven CA predecessors, LLM agents exhibit advanced linguistic capabilities that enable them to produce seemingly coherent and contextually appropriate responses for various tasks. As such, LLMs have quickly been developed into different professional practices of educators, programmers, researchers, engineers, artists, lawyers and even doctors (Eloundou et al. , 2023) . OpenAI reported that approximately 19% of workers, from 80% of the workforce, at least 50% of their tasks impacted by LLM agents (Eloundou et al. , 2023) . The novel, heuristic-probabilistic approach of LLMs to text generation, despite its strengths, however. comes with a price; LLMs are associated with risks and obstacles such as inconsistencies, hallucinations, and misalignment with pragmatic context (Huang et al. , 2025; Banerjee et al. , 2024) . Such risks are inextricably tied to the generativ

FAQs

Why do users need to repair LLM outputs in human-agent conversations?

LLM responses can be unreliable, ambiguous, or misleading, creating interactive trouble that users work to resolve during the exchange.

What data does the paper analyze to study repair work?

It analyzes 21 real-world chat transcripts drawn from diverse work contexts to observe how repairs unfold in practice.

What does the proposed taxonomy cover?

The taxonomy covers different repair initiators, repair stages (trouble classification, trouble specification, trouble management), and repair processes (incremental, grounding, and validating).

Sep 11, 2026

Uploaded by Angel