Beyond Benchmarks: Human-Aligned Evaluation Frameworks for Large Language Models

Rapid advancement of large language models (LLMs), such as OpenAI’s GPT-4, has transformed natural language processing (NLP) in a wide range of applications. However, evaluating these models remains a fundamental challenge, especially in tasks that involve nuanced, subjective human judgment. Traditional automated metrics like BLEU and ROUGE often fail to capture semantic nuance or cultural fidelity, while more recent approaches, such as reinforcement learning from human feedback (RLHF), are data intensive and limited in scope, often restricted to domains like summarization or dialogue. These limitations are particularly acute in open-ended, real-world tasks where human satisfaction and context sensitivity are crucial. This thesis proposes a unified research agenda to rethink how LLMs are evaluated, focusing on three interconnected contributions. First, we introduce a formal framework for analyzing LLM ranking algorithms, identifying desirable properties such as transitivity, prediction accuracy, and robustness to sparse or noisy comparisons. Through empirical analysis, we demonstrate how popular methods such as Elo and Glicko can lead to unstable or misleading rankings in practice. Second, we present SLMEval, a scalable evaluation framework that uses entropy-based calibration to align model-generated scores with human preferences, even when only limited annotated data are available. SLMEval is evaluated across use cases such as Peptalk generation and productivity assistance, where it achieves stronger alignment with human judgments than existing evaluators like GEval and GPTScore. Finally, we contribute GuyLingo, the first large-scale corpus for Guyanese Creole, and use it to explore the challenges of evaluation in low-resource, culturally rich linguistic settings, where both translation accuracy and cultural grounding are essential. Together, these contributions offer a cohesive and practical response to the limitations of current LLM evaluation practices. By addressing algorithmic reliability, scalable preference alignment, and multilingual inclusivity, this thesis aims to bridge the gap between automated metrics and human-centered evaluation, paving the way for a more trustworthy and context-aware deployment of LLMs.