Optimizing Large Language Models from a Data SystemsPerspective

Strong retrieval and reasoning capabilities are essential for large language models (LLMs) to effectively handle a broad spectrum of downstream tasks, such as open-domain question answering and solving math or science problems. While current LLM-based frameworks achieve strong performance on complex retrieval and reasoning tasks, they do so at a high computational cost. Additionally, they often lack structured, systematic problem-solving strategies, leading to unexpected failures. In particular, these models typically operate in an iterative, online, and isolated fashion—failing to exploit relationships across data sources, opportunities for offline computation, and the benefits of reusability—resulting in less-than-optimal outcomes. In contrast, traditional data management systems are engineered for both efficiency and accuracy, with careful coordination across all stages of the query pipeline. Inspired by these principles, this work proposes novel approaches to improve LLMbased retrieval and reasoning by incorporating optimization techniques from data systems. Our evaluation across a range of knowledge- and reasoning-intensive datasets demonstrates significant gains in both accuracy and computational efficiency.

Name thesis:degree_name

Master

Department dc:contributor.department

Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science

Grantor dc:publisher

Massachusetts Institute of Technology

Year dc:date.issued

2025

Author dc:creator

• Chen, Peter Baile

Advisor dc:contributor.advisor

• Cafarella, Michael J.

Statement dc:rights

• In Copyright - Educational Use Permitted
• Copyright retained by author(s)

Licence dc:rights.uri

• https://rightsstatements.org/page/InC-EDU/1.0/

Handle dc:identifier.uri

https://hdl.handle.net/1721.1/164667

OAI identifier oai:identifier

oai:dspace.mit.edu:1721.1/164667