Toward Deliberative AI: Multi-Agent LLMs for Real-World Reasoning
dc.contributor.author | Pitre, Priya Nitin | en |
dc.contributor.committeechair | Wang, Xuan | en |
dc.contributor.committeecochair | Ramakrishnan, Naren | en |
dc.contributor.committeemember | Yanardag, Pinar | en |
dc.contributor.department | Computer Science & Applications | en |
dc.date.accessioned | 2025-06-05T08:01:52Z | en |
dc.date.available | 2025-06-05T08:01:52Z | en |
dc.date.issued | 2025-06-04 | en |
dc.description.abstract | Multi-agent debate has emerged as a promising strategy for improving the reasoning abilities of large language models (LLMs). However, existing approaches often fall short due to inefficiencies, shallow agreement, and a lack of real-world applicability. In this thesis, we introduce two novel frameworks CONSENSAGENT and CCAGENTdesigned to improve both the effectiveness and efficiency of LLM debates across objective and real-world tasks. CONSENSAGENT tackles key limitations such as sycophancy (models blindly agreeing with each other) and ambiguous prompts by introducing a trigger-based architecture that automatically refines prompts using past agent dis- cussions. This results in better reasoning, fewer debate rounds, and reduced computational cost. We evaluate the framework across six benchmark datasets and show that CONSENSAGENT con- sistently outperforms baselines. CCAGENT extends this work to real-world decision-making. We introduce two new datasetsone from interviews with city planners, another from U.S. Senate voting recordsand propose structured debate strategies (e.g., moderation, nudging) along with behavioral metrics (e.g., sycophancy, vote switching). A lightweight few-shot DPO training method is used to align agent behavior with collaborative reasoning goals. Together, these contributions demonstrate how we can move from toy benchmarks to deliberative, scalable systems that better reflect how human decision-making worksand how AI can meaningfully assist it. | en |
dc.description.abstractgeneral | Artificial intelligence (AI) systems are increasingly used to assist with decision-making, yet most current models operate in isolation, producing answers without engaging in deliberation. This the- sis explores a new approach: enabling multiple AI agents to debate, reason together, and reach consensussimilar to how groups of people discuss and resolve complex issues. We present two frameworks that advance this vision. The first, CONSENSAGENT, focuses on tasks with objec- tive answers such as arithmetic and question answering. It addresses key limitations in multi-agent discussions, including inefficiency and a tendency for agents to agree uncritically with one another. By identifying when a conversation is stalled or biased, the system can automatically rewrite the question to clarify ambiguity and encourage meaningful reasoning. This leads to faster and more accurate outcomes.The second framework, CCAGENT, extends this capability to real-world do- mains such as city planning and political policy. It introduces new datasets, structured strategies for guiding debates, and novel evaluation metrics that capture qualities like flexibility, compromise, and alignment. Using a lightweight training method, we demonstrate that AI agents can improve not only their answers, but also the quality of their interactions.Together, these contributions move AI toward more collaborative, transparent, and human-aligned decision-makingoffering a founda- tion for systems that can support complex societal deliberation in a principled and scalable way. | en |
dc.format.medium | ETD | en |
dc.identifier.other | vt_gsexam:44004 | en |
dc.identifier.uri | https://hdl.handle.net/10919/135063 | en |
dc.language.iso | en | en |
dc.publisher | Virginia Tech | en |
dc.rights | In Copyright | en |
dc.rights.uri | http://rightsstatements.org/vocab/InC/1.0/ | en |
dc.subject | Transformer Models | en |
dc.subject | LLMs | en |
dc.subject | Complex Reasoning | en |
dc.subject | Multi-Agent Architectures | en |
dc.title | Toward Deliberative AI: Multi-Agent LLMs for Real-World Reasoning | en |
dc.type | Thesis | en |
thesis.degree.grantor | Virginia Polytechnic Institute and State University | en |
thesis.degree.level | masters | en |
Files
Original bundle
1 - 1 of 1