본문

서브메뉴

Human-AI Interaction in the Era of Large Language Models (LLMs)
Human-AI Interaction in the Era of Large Language Models (LLMs)
Human-AI Interaction in the Era of Large Language Models (LLMs)

상세정보

자료유형  
 학위논문 서양
최종처리일시  
20260202103614
ISBN  
9798290943916
DDC  
004
저자명  
Mohammadi, Behnam.
서명/저자  
Human-AI Interaction in the Era of Large Language Models (LLMs)
발행사항  
[Sl] : Carnegie Mellon University, 2025
발행사항  
Ann Arbor : ProQuest Dissertations & Theses, 2025
형태사항  
141 p
주기사항  
Source: Dissertations Abstracts International, Volume: 87-02, Section: B.
주기사항  
Advisor: Derdenger, Tim.
학위논문주기  
Thesis (Ph.D.)--Carnegie Mellon University, 2025.
초록/해제  
요약The first chapter, joint work with Nikhil Malik, Tim Derdenger, and Kannan Srinivasan, challenges conventional wisdom regarding eXplainable AI (XAI) regulations such as GDPR. Through a game-theoretic model examining XAI methods and levels in a duopoly market with heterogeneous customer preferences, we demonstrate that partial explanations can emerge as an equilibrium in unregulated settings. Importantly, we identify conditions where mandating full explanations through regulation may actually harm consumer surplus rather than enhance it. This finding holds across various policy levers (strict, self-regulating, and lower bound), regardless of firms' choice of XAI methods and policy objectives including welfare maximization, consumer surplus, and average XAI depth. Our comparative analysis reveals that while strict XAI policies ensure uniform explanation depth, they potentially limit firms' capacity for differentiation and innovation. Conversely, unregulated XAI, while offering maximum flexibility, may fail to guarantee minimum explanation depth for all consumers. The introduction of flexible approaches-self-regulating XAI and lower-bounded XAI-results in higher consumer welfare than either unregulated or full XAI policies. This research urges policymakers to consider a more nuanced approach when crafting XAI regulations, as a one-size-fits-all policy across all markets, particularly one mandating full explanation, may not yield the desired outcomes. For firms operating in these markets, the optimal strategy may not be to provide full explanations, as partial explanations can emerge as equilibrium strategies that better serve their competitive positioning while still addressing consumer needs.The second chapter addresses the growing use of LLMs as simulated consumers in marketing research. I develop a novel approach based on Shapley values from cooperative game theory to interpret LLM behavior and quantify the relative contribution of prompt components to model outputs. Through applications in discrete choice experiments and cognitive bias investigations, I uncover what I term the "token noise" effect-a phenomenon where LLM decisions are disproportionately influenced by tokens providing minimal informative content (such as empty lines in a questionnaire!). This finding provides a theoretical foundation for understanding how LLMs process information and make decisions, revealing fundamental differences from human cognition that must be accounted for in marketing research. For marketers employing LLMs for consumer simulation, this raises significant concerns about the validity of using LLMs as proxies for human subjects and necessitates rigorous validation procedures when using LLMs for preference elicitation or behavior prediction. The proposed Shapley value method offers practitioners a model-agnostic approach for optimizing prompts and mitigating apparent cognitive biases in LLM responses.The third chapter investigates the unintended consequences of AI alignment techniques on the creative capabilities of language models. Through a series of experiments with the Llama model family (created by Meta/Facebook), I demonstrate that alignment methods like Reinforcement Learning from Human Feedback (RLHF), while reducing bias and harmful outputs, significantly diminish syntactic and semantic diversity. My findings reveal that aligned models exhibit lower entropy in token predictions, form distinct clusters in embedding space, and gravitate toward "attractor states", indicating limited output diversity. This contributes to our theoretical understanding of AI creativity by conceptualizing the relationship between alignment and creativity as a fundamental trade-off rather than a technical limitation. Marketing teams must strategically balance the benefits of AI safety alignment with creative performance when selecting language models for content generation tasks. Different models may be optimal for different marketing functions-aligned models for customer-facing interactions where consistency and brand safety are paramount, and base models for ideation tasks that benefit from novelty and creativity, such as ad copywriting and customer persona development.The fourth chapter steps beyond individual models to examine networks of AI agents that work together to accomplish complex goals, such as automating various functions in a business (e.g., customer support, SEO, refunds, etc.). This introduces a new challenge: not just how we build these agents, but how we coordinate them. To address this, I introduce Pel, a programming language I developed from scratch specifically for orchestrating AI agents. Pel offers an elegant, principled framework for multi-agent AI systems, addressing limitations in current methods of controlling LLMs through a syntactically simple yet semantically rich platform for expressing complex actions, control flow, and inter-agent communication. Its design emphasizes minimal grammar suitable for constrained LLM generation, powerful composition mechanisms, and built-in support for natural language conditions. This advances programming language theory through the development of a domain-specific language (DSL) optimized for AI agent control, proposing a new paradigm for human-AI interaction that incorporates the unique capabilities and limitations of language models. From a managerial perspective, Pel provides marketing technology teams with a specialized tool for building sophisticated marketing automation systems powered by LLMs, to be used in customer engagement and support, content personalization, and multi-channel campaign management.
일반주제명  
Computer science
일반주제명  
Engineering
키워드  
EXplainable AI
키워드  
Shapley value
키워드  
Interpretability
키워드  
Large language models
키워드  
Machine learning
키워드  
Cognitive biases
기타저자  
Carnegie Mellon University Tepper School of Business
기본자료저록  
Dissertations Abstracts International. 87-02B.
전자적 위치 및 접속  
로그인 후 원문을 볼 수 있습니다.

MARC

 008260126s2025        us                              c    eng  d
■001000017357893
■00520260202103614
■006m          o    d                
■007cr#unu||||||||
■020    ▼a9798290943916
■035    ▼a(MiAaPQ)AAI32044207
■040    ▼aMiAaPQ▼cMiAaPQ
■0820  ▼a004
■1001  ▼aMohammadi,  Behnam.▼0(orcid)0000-0003-0028-8246
■24510▼aHuman-AI  Interaction  in  the  Era  of  Large  Language  Models  (LLMs)
■260    ▼a[Sl]▼bCarnegie  Mellon  University▼c2025
■260  1▼aAnn  Arbor▼bProQuest  Dissertations  &  Theses▼c2025
■300    ▼a141  p
■500    ▼aSource:  Dissertations  Abstracts  International,  Volume:  87-02,  Section:  B.
■500    ▼aAdvisor:  Derdenger,  Tim.
■5021  ▼aThesis  (Ph.D.)--Carnegie  Mellon  University,  2025.
■520    ▼aThe  first  chapter,  joint  work  with  Nikhil  Malik,  Tim  Derdenger,  and  Kannan  Srinivasan,  challenges  conventional  wisdom  regarding  eXplainable  AI  (XAI)  regulations  such  as  GDPR.  Through  a  game-theoretic  model  examining  XAI  methods  and  levels  in  a  duopoly  market  with  heterogeneous  customer  preferences,  we  demonstrate  that  partial  explanations  can  emerge  as  an  equilibrium  in  unregulated  settings.  Importantly,  we  identify  conditions  where  mandating  full  explanations  through  regulation  may  actually  harm  consumer  surplus  rather  than  enhance  it.  This  finding  holds  across  various  policy  levers  (strict,  self-regulating,  and  lower  bound),  regardless  of  firms'  choice  of  XAI  methods  and  policy  objectives  including  welfare  maximization,  consumer  surplus,  and  average  XAI  depth.  Our  comparative  analysis  reveals  that  while  strict  XAI  policies  ensure  uniform  explanation  depth,  they  potentially  limit  firms'  capacity  for  differentiation  and  innovation.  Conversely,  unregulated  XAI,  while  offering  maximum  flexibility,  may  fail  to  guarantee  minimum  explanation  depth  for  all  consumers.  The  introduction  of  flexible  approaches-self-regulating  XAI  and  lower-bounded  XAI-results  in  higher  consumer  welfare  than  either  unregulated  or  full  XAI  policies.  This  research  urges  policymakers  to  consider  a  more  nuanced  approach  when  crafting  XAI  regulations,  as  a  one-size-fits-all  policy  across  all  markets,  particularly  one  mandating  full  explanation,  may  not  yield  the  desired  outcomes.  For  firms  operating  in  these  markets,  the  optimal  strategy  may  not  be  to  provide  full  explanations,  as  partial  explanations  can  emerge  as  equilibrium  strategies  that  better  serve  their  competitive  positioning  while  still  addressing  consumer  needs.The  second  chapter  addresses  the  growing  use  of  LLMs  as  simulated  consumers  in  marketing  research.  I  develop  a  novel  approach  based  on  Shapley  values  from  cooperative  game  theory  to  interpret  LLM  behavior  and  quantify  the  relative  contribution  of  prompt  components  to  model  outputs.  Through  applications  in  discrete  choice  experiments  and  cognitive  bias  investigations,  I  uncover  what  I  term  the  "token  noise"  effect-a  phenomenon  where  LLM  decisions  are  disproportionately  influenced  by  tokens  providing  minimal  informative  content  (such  as  empty  lines  in  a  questionnaire!).  This  finding  provides  a  theoretical  foundation  for  understanding  how  LLMs  process  information  and  make  decisions,  revealing  fundamental  differences  from  human  cognition  that  must  be  accounted  for  in  marketing  research.  For  marketers  employing  LLMs  for  consumer  simulation,  this  raises  significant  concerns  about  the  validity  of  using  LLMs  as  proxies  for  human  subjects  and  necessitates  rigorous  validation  procedures  when  using  LLMs  for  preference  elicitation  or  behavior  prediction.  The  proposed  Shapley  value  method  offers  practitioners  a  model-agnostic  approach  for  optimizing  prompts  and  mitigating  apparent  cognitive  biases  in  LLM  responses.The  third  chapter  investigates  the  unintended  consequences  of  AI  alignment  techniques  on  the  creative  capabilities  of  language  models.  Through  a  series  of  experiments  with  the  Llama  model  family  (created  by  Meta/Facebook),  I  demonstrate  that  alignment  methods  like  Reinforcement  Learning  from  Human  Feedback  (RLHF),  while  reducing  bias  and  harmful  outputs,  significantly  diminish  syntactic  and  semantic  diversity.  My  findings  reveal  that  aligned  models  exhibit  lower  entropy  in  token  predictions,  form  distinct  clusters  in  embedding  space,  and  gravitate  toward  "attractor  states",  indicating  limited  output  diversity.  This  contributes  to  our  theoretical  understanding  of  AI  creativity  by  conceptualizing  the  relationship  between  alignment  and  creativity  as  a  fundamental  trade-off  rather  than  a  technical  limitation.  Marketing  teams  must  strategically  balance  the  benefits  of  AI  safety  alignment  with  creative  performance  when  selecting  language  models  for  content  generation  tasks.  Different  models  may  be  optimal  for  different  marketing  functions-aligned  models  for  customer-facing  interactions  where  consistency  and  brand  safety  are  paramount,  and  base  models  for  ideation  tasks  that  benefit  from  novelty  and  creativity,  such  as  ad  copywriting  and  customer  persona  development.The  fourth  chapter  steps  beyond  individual  models  to  examine  networks  of  AI  agents  that  work  together  to  accomplish  complex  goals,  such  as  automating  various  functions  in  a  business  (e.g.,  customer  support,  SEO,  refunds,  etc.).  This  introduces  a  new  challenge:  not  just  how  we  build  these  agents,  but  how  we  coordinate  them.  To  address  this,  I  introduce  Pel,  a  programming  language  I  developed  from  scratch  specifically  for  orchestrating  AI  agents.  Pel  offers  an  elegant,  principled  framework  for  multi-agent  AI  systems,  addressing  limitations  in  current  methods  of  controlling  LLMs  through  a  syntactically  simple  yet  semantically  rich  platform  for  expressing  complex  actions,  control  flow,  and  inter-agent  communication.  Its  design  emphasizes  minimal  grammar  suitable  for  constrained  LLM  generation,  powerful  composition  mechanisms,  and  built-in  support  for  natural  language  conditions.  This  advances  programming  language  theory  through  the  development  of  a  domain-specific  language  (DSL)  optimized  for  AI  agent  control,  proposing  a  new  paradigm  for  human-AI  interaction  that  incorporates  the  unique  capabilities  and  limitations  of  language  models.  From  a  managerial  perspective,  Pel  provides  marketing  technology  teams  with  a  specialized  tool  for  building  sophisticated  marketing  automation  systems  powered  by  LLMs,  to  be  used  in  customer  engagement  and  support,  content  personalization,  and  multi-channel  campaign  management.
■590    ▼aSchool  code:  0041.
■650  4▼aComputer  science
■650  4▼aEngineering
■653    ▼aEXplainable  AI
■653    ▼aShapley  value
■653    ▼aInterpretability
■653    ▼aLarge  language  models
■653    ▼aMachine  learning
■653    ▼aCognitive  biases
■690    ▼a0338
■690    ▼a0800
■690    ▼a0984
■690    ▼a0537
■71020▼aCarnegie  Mellon  University▼bTepper  School  of  Business.
■7730  ▼tDissertations  Abstracts  International▼g87-02B.
■790    ▼a0041
■791    ▼aPh.D.
■792    ▼a2025
■793    ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357893▼nKERIS▼z이  자료의  원문은  한국교육학술정보원에서  제공합니다.

미리보기

내보내기

chatGPT토론

Ai 추천 관련 도서


    신착도서 더보기
    최근 3년간 통계입니다.

    소장정보

    • 예약
    • 소재불명신고
    • 나의폴더
    • 우선정리요청
    • 비도서대출신청
    • 야간 도서대출신청
    소장자료
    등록번호 청구기호 소장처 대출가능여부 대출정보
    TF15177 전자도서 대출가능 마이폴더 부재도서신고 비도서대출신청 야간 도서대출신청

    * 대출중인 자료에 한하여 예약이 가능합니다. 예약을 원하시면 예약버튼을 클릭하십시오.

    해당 도서를 다른 이용자가 함께 대출한 도서

    관련 인기도서

    로그인 후 이용 가능합니다.