Detect, Explain, Ground: A Sustainable Pipeline for Robust Large Language Model–Powered Agents

Large language models (LLMs) enable fluent dialogue but still suffer from critical breakdowns, such as incoherence, irrelevance, or factual inaccuracies (hallucinations). This thesis develops a sustainable, three-stage pipeline to detect, manage, and ground LLM-powered agents, enhancing their robustness and reliability. First, to establish a baseline for self-monitoring, we evaluate general-purpose LLMs on DBDC5, demonstrating that models like GPT-4 can surpass specialized detectors. While effective, relying on such large models for continuous monitoring is computationally unsustainable. Building on this finding, we propose a novel "monitor-explain-escalate" architecture. In this system, a lightweight detector provides real-time analysis and escalates high-risk conversational turns to a stronger model, preserving quality while significantly reducing costs. Finally, to address the core issue of factual accuracy that underlies many breakdowns, we introduce a Hierarchical Lexical Graph to improve evidence retrieval for multi-hop question answering. Across five datasets, this graph-augmented retrieval method achieves a 23.1% relative improvement in recall and correctness over baseline RAG systems. Together, these contributions advance the development of reliable, interpretable, and resource-efficient LLM agents by establishing a comprehensive framework that manages conversational flow, operationalizes low-carbon management, and grounds multi-document reasoning in verifiable evidence.

Department dc:contributor.department

Electrical and Computer Engineering

Year dc:date.issued

2025

Author dc:creator

• Ghassel, Abdellah

Advisor dc:contributor.supervisor

• Zhu, Xiaodan

• conversational artificial intelligence
• graph machine learning
• large language models

Statement dc:rights

• Attribution-ShareAlike 4.0 International

Licence dc:rights.uri

• http://creativecommons.org/licenses/by-sa/4.0/

Language dc:language.iso

eng

Handle dc:identifier.uri

https://hdl.handle.net/1974/35343

OAI identifier oai:identifier

oai:queensu.scholaris.ca:1974/35343