Enhancing large language models: toward more reliable and equitable NLP

Recent advances in large language models (LLMs) have enabled them to achieve striking fewshot performance on complex tasks, yet they can remain sensitive to prompt configurations and prone to subtle biases. This thesis addresses two central problems: (1) selecting informative few-shot examples and (2) localizing and mitigating bias in ambiguous comparative prompts. First, we propose a complexity-based approach for selecting examples in few-shot sequence tagging tasks, aiming to align test examples with training examples based on syntactic and semantic metrics. By focusing on features like sentence similarity, length matching, and label diversity, we achieve more consistent and robust outcomes without fine-tuning or adding parameters. Second, we develop a method to localize and mitigate bias in LLMs by examining the attention layers that favor certain entities over others. We then scale attention in those identified layers to reduce skewed preferences, preserving overall model fluency while mitigating biases. Extensive evaluations confirm that this targeted attention manipulation provides a lightweight way to address fairness concerns without sacrificing downstream accuracy.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Adiga, Rishabh

Contributors dc:contributor

• Chandrasekaran, Varun

• LLM
• Few-shot
• Attention
• Bias
• Prompting
• Fairness
• Metrics
• Localization
• Mitigation

Statement dc:rights

• © 2025 Rishabh Adiga

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129259

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129259