LARGE LANGUAGE MODELS (LLMS) ARE TRANSFORMING AI APPLICATIONS, DEMONSTRATING REMARKABLE ABILITIES IN NATURAL LANGUAGE TASKS. THEIR DEPLOYMENT IN SOFTWARE ENGINEERING AND HEALTHCARE OFFERS EXCITING OPPORTUNITIES, AS THE TEXTUAL NATURE OF THESE DOMAINS MAKES THEM IDEAL FOR LLM PROCESSING. HOWEVER, THESE APPLICATIONS MUST MEET STRINGENT REQUIREMENTS FOR ACCURACY, CONSISTENCY, AND SAFETY WHILE INTEGRATING INTO PROFESSIONAL WORKFLOWS. DESPITE GROWING RESEARCH INTEREST EXPLORING DIVERSE ARCHITECTURES AND ADAPTATION METHODS, SIGNIFICANT CHALLENGES THREATEN REAL-WORLD PRACTICALITY. EVIDENCE REMAINS FRAGMENTED REGARDING: (I) ADAPTATION STRATEGY PERFORMANCE ACROSS SPECIALIZED DOMAINS, (II) MODEL ARCHITECTURES' TASK-SPECIFIC STRENGTHS AND WEAKNESSES, AND (III) RELIABILITY REQUIREMENTS FOR HIGH-STAKES ENVIRONMENTS. THIS THESIS ADDRESSES THESE GAPS BY DEVELOPING NOVEL LLM ADAPTATION AND DEPLOYMENT APPROACHES FOR SOFTWARE ENGINEERING AND HEALTHCARE. WE PRESENT CONTRIBUTIONS INCLUDING CONFIDENCE-BASED FRAMEWORKS, MODULAR ARCHITECTURES, AND DOMAIN-SPECIFIC METHODS, EVALUATED ACROSS APPLICATIONS LIKE USE CASE IMPROVEMENT, AUTOMATED TESTING, CLINICAL DECISION SUPPORT, AND DRUG INTERACTION PREDICTION. OUR FINDINGS CHALLENGE COMMON ASSUMPTIONS. FIRST, MODEL SIZE DOESN'T CORRELATE DIRECTLY WITH EFFECTIVENESS—APPROPRIATELY ADAPTED SMALL MODELS MATCH LARGER SYSTEMS WHILE ENABLING LOCAL DEPLOYMENT WITH ENHANCED DATA SOVEREIGNTY. SECOND, ADAPTATION STRATEGIES PRODUCE DRAMATICALLY DIFFERENT OUTCOMES ACROSS LLMS, REVEALING METHOD-MODEL INTERACTION COMPLEXITIES THAT RESIST UNIVERSAL SOLUTIONS. THIRD, CONFIDENCE-BASED FRAMEWORKS SIMULTANEOUSLY ENHANCE SAFETY AND EFFECTIVENESS, CONTRARY TO EXPECTATIONS OF INHERENT TRADE-OFFS. USER ACCEPTANCE DEPENDS ON DECISION-MAKING TRANSPARENCY AND AUTONOMY, NOT JUST RAW PERFORMANCE. WE IDENTIFY CRITICAL OPEN RESEARCH DIRECTIONS: PREDICTIVE MODEL-METHOD COMPATIBILITY FRAMEWORKS, DYNAMIC CONFIDENCE CALIBRATION, SYSTEMATIC BIAS DETECTION IN ENSEMBLE FAILURES, AND LONGITUDINAL STABILITY MECHANISMS. THESE CONTRIBUTIONS ADVANCE UNDERSTANDING OF EFFECTIVE LLM DEPLOYMENT IN SPECIALIZED, HIGH-STAKES DOMAINS.
LARGE LANGUAGE MODELS (LLMS) ARE TRANSFORMING AI APPLICATIONS, DEMONSTRATING REMARKABLE ABILITIES IN NATURAL LANGUAGE TASKS. THEIR DEPLOYMENT IN SOFTWARE ENGINEERING AND HEALTHCARE OFFERS EXCITING OPPORTUNITIES, AS THE TEXTUAL NATURE OF THESE DOMAINS MAKES THEM IDEAL FOR LLM PROCESSING. HOWEVER, THESE APPLICATIONS MUST MEET STRINGENT REQUIREMENTS FOR ACCURACY, CONSISTENCY, AND SAFETY WHILE INTEGRATING INTO PROFESSIONAL WORKFLOWS. DESPITE GROWING RESEARCH INTEREST EXPLORING DIVERSE ARCHITECTURES AND ADAPTATION METHODS, SIGNIFICANT CHALLENGES THREATEN REAL-WORLD PRACTICALITY. EVIDENCE REMAINS FRAGMENTED REGARDING: (I) ADAPTATION STRATEGY PERFORMANCE ACROSS SPECIALIZED DOMAINS, (II) MODEL ARCHITECTURES' TASK-SPECIFIC STRENGTHS AND WEAKNESSES, AND (III) RELIABILITY REQUIREMENTS FOR HIGH-STAKES ENVIRONMENTS. THIS THESIS ADDRESSES THESE GAPS BY DEVELOPING NOVEL LLM ADAPTATION AND DEPLOYMENT APPROACHES FOR SOFTWARE ENGINEERING AND HEALTHCARE. WE PRESENT CONTRIBUTIONS INCLUDING CONFIDENCE-BASED FRAMEWORKS, MODULAR ARCHITECTURES, AND DOMAIN-SPECIFIC METHODS, EVALUATED ACROSS APPLICATIONS LIKE USE CASE IMPROVEMENT, AUTOMATED TESTING, CLINICAL DECISION SUPPORT, AND DRUG INTERACTION PREDICTION. OUR FINDINGS CHALLENGE COMMON ASSUMPTIONS. FIRST, MODEL SIZE DOESN'T CORRELATE DIRECTLY WITH EFFECTIVENESS—APPROPRIATELY ADAPTED SMALL MODELS MATCH LARGER SYSTEMS WHILE ENABLING LOCAL DEPLOYMENT WITH ENHANCED DATA SOVEREIGNTY. SECOND, ADAPTATION STRATEGIES PRODUCE DRAMATICALLY DIFFERENT OUTCOMES ACROSS LLMS, REVEALING METHOD-MODEL INTERACTION COMPLEXITIES THAT RESIST UNIVERSAL SOLUTIONS. THIRD, CONFIDENCE-BASED FRAMEWORKS SIMULTANEOUSLY ENHANCE SAFETY AND EFFECTIVENESS, CONTRARY TO EXPECTATIONS OF INHERENT TRADE-OFFS. USER ACCEPTANCE DEPENDS ON DECISION-MAKING TRANSPARENCY AND AUTONOMY, NOT JUST RAW PERFORMANCE. WE IDENTIFY CRITICAL OPEN RESEARCH DIRECTIONS: PREDICTIVE MODEL-METHOD COMPATIBILITY FRAMEWORKS, DYNAMIC CONFIDENCE CALIBRATION, SYSTEMATIC BIAS DETECTION IN ENSEMBLE FAILURES, AND LONGITUDINAL STABILITY MECHANISMS. THESE CONTRIBUTIONS ADVANCE UNDERSTANDING OF EFFECTIVE LLM DEPLOYMENT IN SPECIALIZED, HIGH-STAKES DOMAINS.
LEVERAGING LLMS FOR SOFTWARE ENGINEERING AND HEALTHCARE: ADAPTATION STRATEGIES AND EMPIRICAL EVALUATION / Gabriele De Vito , 2025 Dec 05. 38. ciclo, Anno Accademico 2024/25.
LEVERAGING LLMS FOR SOFTWARE ENGINEERING AND HEALTHCARE: ADAPTATION STRATEGIES AND EMPIRICAL EVALUATION
DE VITO, GABRIELE
2025
Abstract
LARGE LANGUAGE MODELS (LLMS) ARE TRANSFORMING AI APPLICATIONS, DEMONSTRATING REMARKABLE ABILITIES IN NATURAL LANGUAGE TASKS. THEIR DEPLOYMENT IN SOFTWARE ENGINEERING AND HEALTHCARE OFFERS EXCITING OPPORTUNITIES, AS THE TEXTUAL NATURE OF THESE DOMAINS MAKES THEM IDEAL FOR LLM PROCESSING. HOWEVER, THESE APPLICATIONS MUST MEET STRINGENT REQUIREMENTS FOR ACCURACY, CONSISTENCY, AND SAFETY WHILE INTEGRATING INTO PROFESSIONAL WORKFLOWS. DESPITE GROWING RESEARCH INTEREST EXPLORING DIVERSE ARCHITECTURES AND ADAPTATION METHODS, SIGNIFICANT CHALLENGES THREATEN REAL-WORLD PRACTICALITY. EVIDENCE REMAINS FRAGMENTED REGARDING: (I) ADAPTATION STRATEGY PERFORMANCE ACROSS SPECIALIZED DOMAINS, (II) MODEL ARCHITECTURES' TASK-SPECIFIC STRENGTHS AND WEAKNESSES, AND (III) RELIABILITY REQUIREMENTS FOR HIGH-STAKES ENVIRONMENTS. THIS THESIS ADDRESSES THESE GAPS BY DEVELOPING NOVEL LLM ADAPTATION AND DEPLOYMENT APPROACHES FOR SOFTWARE ENGINEERING AND HEALTHCARE. WE PRESENT CONTRIBUTIONS INCLUDING CONFIDENCE-BASED FRAMEWORKS, MODULAR ARCHITECTURES, AND DOMAIN-SPECIFIC METHODS, EVALUATED ACROSS APPLICATIONS LIKE USE CASE IMPROVEMENT, AUTOMATED TESTING, CLINICAL DECISION SUPPORT, AND DRUG INTERACTION PREDICTION. OUR FINDINGS CHALLENGE COMMON ASSUMPTIONS. FIRST, MODEL SIZE DOESN'T CORRELATE DIRECTLY WITH EFFECTIVENESS—APPROPRIATELY ADAPTED SMALL MODELS MATCH LARGER SYSTEMS WHILE ENABLING LOCAL DEPLOYMENT WITH ENHANCED DATA SOVEREIGNTY. SECOND, ADAPTATION STRATEGIES PRODUCE DRAMATICALLY DIFFERENT OUTCOMES ACROSS LLMS, REVEALING METHOD-MODEL INTERACTION COMPLEXITIES THAT RESIST UNIVERSAL SOLUTIONS. THIRD, CONFIDENCE-BASED FRAMEWORKS SIMULTANEOUSLY ENHANCE SAFETY AND EFFECTIVENESS, CONTRARY TO EXPECTATIONS OF INHERENT TRADE-OFFS. USER ACCEPTANCE DEPENDS ON DECISION-MAKING TRANSPARENCY AND AUTONOMY, NOT JUST RAW PERFORMANCE. WE IDENTIFY CRITICAL OPEN RESEARCH DIRECTIONS: PREDICTIVE MODEL-METHOD COMPATIBILITY FRAMEWORKS, DYNAMIC CONFIDENCE CALIBRATION, SYSTEMATIC BIAS DETECTION IN ENSEMBLE FAILURES, AND LONGITUDINAL STABILITY MECHANISMS. THESE CONTRIBUTIONS ADVANCE UNDERSTANDING OF EFFECTIVE LLM DEPLOYMENT IN SPECIALIZED, HIGH-STAKES DOMAINS.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi Elettronica.pdf
accesso aperto
Descrizione: Tesi Elettronica
Tipologia:
Tesi di dottorato
Dimensione
11.26 MB
Formato
Adobe PDF
|
11.26 MB | Adobe PDF | Visualizza/Apri |
|
Abstract.pdf
accesso aperto
Descrizione: Abstract
Tipologia:
Tesi di dottorato
Dimensione
85.12 kB
Formato
Adobe PDF
|
85.12 kB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


