{"id":{"repo_id":"uiuc","oai_identifier":"oai:www.ideals.illinois.edu:2142/132559"},"canonical_url":"https://search.dev.ndltd.org/etd/uiuc/oai:www.ideals.illinois.edu:2142/132559","repository":{"repo_id":"uiuc","name":"University of Illinois - Urbana-Champaign","base_url":"https://www.ideals.illinois.edu/oai-pmh"},"display":{"title":"Malice, inequality, instability, or ignorance? Disentangling the mechanisms of LLM unfairness","abstract":"Ensuring fairness in large language models (LLMs) is critical as these models are increasingly deployed in sensitive domains. Traditional fairness metrics typically report a single scalar score, which conflates distinct sources of model failure and obscures underlying biases. In this work, we propose a Hierarchical Bias-Variance Decomposition framework—termed BDSU—that decomposes total discrimination risk into four interpretable components: Bias (systematic global error), Disparity (group-level variance), Sensitivity (context-level variance), and Uncertainty (stochastic or token-level variance). By applying the law of total variance recursively, BDSU provides a principled method to quantify and separate these failure modes, aligning each with ethical and reliability priorities. We further introduce a conditional micro-diagnosis to evaluate fairness at the group level, enabling fine-grained auditing and targeted interventions. Our theoretical framework lays the foundation for more transparent, actionable, and robust evaluation of LLM fairness, highlighting the distinct mechanisms by which models may perpetuate bias or exhibit instability.","abstract_html":"Ensuring fairness in large language models (LLMs) is critical as these models are increasingly deployed in sensitive domains. Traditional fairness metrics typically report a single scalar score, which conflates distinct sources of model failure and obscures underlying biases. In this work, we propose a Hierarchical Bias-Variance Decomposition framework—termed BDSU—that decomposes total discrimination risk into four interpretable components: Bias (systematic global error), Disparity (group-level variance), Sensitivity (context-level variance), and Uncertainty (stochastic or token-level variance). By applying the law of total variance recursively, BDSU provides a principled method to quantify and separate these failure modes, aligning each with ethical and reliability priorities. We further introduce a conditional micro-diagnosis to evaluate fairness at the group level, enabling fine-grained auditing and targeted interventions. Our theoretical framework lays the foundation for more transparent, actionable, and robust evaluation of LLM fairness, highlighting the distinct mechanisms by which models may perpetuate bias or exhibit instability.","abstract_has_math":false,"creators":["Yang, Ke"],"institution":"University of Illinois Urbana-Champaign","degree_name":"M.S.","degree_level":"Thesis","degree_discipline":"Computer Science","degree_department":null,"school":null,"contributors":["Zhai, ChengXiang"],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2025,"date_issued":"2025-12","date_published":"2025-12","updated_at":"2026-07-22T22:25:07Z","subjects":["large language model","unfairness measurement"],"languages":["en"],"rights":["Copyright © 2025 Ke Yang. All rights reserved."],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"https://hdl.handle.net/2142/132559","outbound_label":"Handle","outbound_source":"dc:identifier"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Zhai, ChengXiang"]},{"key":"dc:creator","label":"Author","values":["Yang, Ke"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2025-12","2025-12-03"]},{"key":"dc:type","label":"Dc Type","values":["text","Thesis"]},{"key":"thesis:degree_discipline","label":"Discipline","values":["Computer Science"]},{"key":"thesis:degree_level","label":"Degree Level","values":["Thesis"]},{"key":"thesis:degree_name","label":"Degree Name","values":["M.S."]},{"key":"thesis:institution_name","label":"Thesis Institution Name","values":["University of Illinois Urbana-Champaign"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["large language model","unfairness measurement"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["en"]},{"key":"dc:rights","label":"Dc Rights","values":["Copyright © 2025 Ke Yang. All rights reserved."]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["https://hdl.handle.net/2142/132559"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Ensuring fairness in large language models (LLMs) is critical as these models are increasingly deployed in sensitive domains. Traditional fairness metrics typically report a single scalar score, which conflates distinct sources of model failure and obscures underlying biases. In this work, we propose a Hierarchical Bias-Variance Decomposition framework—termed BDSU—that decomposes total discrimination risk into four interpretable components: Bias (systematic global error), Disparity (group-level variance), Sensitivity (context-level variance), and Uncertainty (stochastic or token-level variance). By applying the law of total variance recursively, BDSU provides a principled method to quantify and separate these failure modes, aligning each with ethical and reliability priorities. We further introduce a conditional micro-diagnosis to evaluate fairness at the group level, enabling fine-grained auditing and targeted interventions. Our theoretical framework lays the foundation for more transparent, actionable, and robust evaluation of LLM fairness, highlighting the distinct mechanisms by which models may perpetuate bias or exhibit instability.","Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2026-02-19 without embargo terms","The student, Ke Yang, accepted the attached license on 2025-12-02 at 20:58.","The student, Ke Yang, submitted this Thesis for approval on 2025-12-03 at 09:48.","This Thesis was approved for publication on 2025-12-03 at 14:34.","DSpace SAF Submission Ingestion Package generated from Vireo submission #23031 on 2026-02-19 at 18:26:20"]},{"key":"dc:format","label":"Dc Format","values":["application/pdf"]},{"key":"dc:title","label":"Title","values":["Malice, inequality, instability, or ignorance? Disentangling the mechanisms of LLM unfairness"]}]}],"canonical_facts":{"dc:contributor":["Zhai, ChengXiang"],"dc:creator":["Yang, Ke"],"dc:date":["2025-12","2025-12-03"],"dc:description":["Ensuring fairness in large language models (LLMs) is critical as these models are increasingly deployed in sensitive domains. Traditional fairness metrics typically report a single scalar score, which conflates distinct sources of model failure and obscures underlying biases. In this work, we propose a Hierarchical Bias-Variance Decomposition framework—termed BDSU—that decomposes total discrimination risk into four interpretable components: Bias (systematic global error), Disparity (group-level variance), Sensitivity (context-level variance), and Uncertainty (stochastic or token-level variance). By applying the law of total variance recursively, BDSU provides a principled method to quantify and separate these failure modes, aligning each with ethical and reliability priorities. We further introduce a conditional micro-diagnosis to evaluate fairness at the group level, enabling fine-grained auditing and targeted interventions. Our theoretical framework lays the foundation for more transparent, actionable, and robust evaluation of LLM fairness, highlighting the distinct mechanisms by which models may perpetuate bias or exhibit instability.","Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2026-02-19 without embargo terms","The student, Ke Yang, accepted the attached license on 2025-12-02 at 20:58.","The student, Ke Yang, submitted this Thesis for approval on 2025-12-03 at 09:48.","This Thesis was approved for publication on 2025-12-03 at 14:34.","DSpace SAF Submission Ingestion Package generated from Vireo submission #23031 on 2026-02-19 at 18:26:20"],"dc:format":["application/pdf"],"dc:identifier":["https://hdl.handle.net/2142/132559"],"dc:language":["en"],"dc:rights":["Copyright © 2025 Ke Yang. All rights reserved."],"dc:subject":["large language model","unfairness measurement"],"dc:title":["Malice, inequality, instability, or ignorance? Disentangling the mechanisms of LLM unfairness"],"dc:type":["text","Thesis"],"thesis:degree_discipline":["Computer Science"],"thesis:degree_level":["Thesis"],"thesis:degree_name":["M.S."],"thesis:institution_name":["University of Illinois Urbana-Champaign"]},"updated_at":"2026-07-22T22:25:07Z"}