Adaptive Dialogue Repair in Conversational Agents with Large Language Models

Abstract:

Conversational agents are widely applied in domains such as education, customer support, or healthcare. Traditional systems rely on natural language understanding pipelines trained for restricted intent spaces. While effective in-domain, these pipelines struggle with out-of-distribution input, noisy utterances, or ambiguity, often leading to conversational failures.

Large language models exhibit strong capabilities for processing unanticipated or noisy input and generating context-sensitive clarifications. This makes them promising for dialogue repair, restoring dialogue interaction flows when misunderstandings occur. Yet, their deployment is limited by computational cost, latency, and the need for consistent and adaptive behavior. This thesis explores the integration of LLM-based dialogue repair mechanisms into dialogue management frameworks. The focus lies on balancing robustness with efficiency and on assessing user perceptions. This thesis will be conducted in the context of the AssistD research project in collaboration with the company ALMA PHIL.
 

Research Questions:

RQ1: How can LLM-based dialogue repair strategies manage out-of-distribution or ambiguous input in dialogue systems?

RQ2: Which optimization techniques can reduce cost and latency of repair while maintaining contextual accuracy?

RQ3: How do users perceive the LLM-driven repair, and how can feedback guide iterative improvements?