Performance and Efficiency Trade-Offs in Fine-Tuned and Pruned Large Language Models

Large Language Models (LLMs) achieve strong performance across natural language processing and code-generation tasks, but their substantial computational and memory requirements limit deployment in resource-constrained environments. Compression techniques such as fine-tuning, and pruning can improve efficiency, yet they are typically evaluated using deterministic metrics that do not adequately capture stochastic variability or reliability degradation. As a result, the practical trade-offs among performance, stability, and efficiency under compression remain insufficiently understood. This study presents a reliability-aware empirical framework for analyzing LLM compression by jointly evaluating performance, stability, and efficiency across varying pruning levels. A LLaMA-based model was fine-tuned using the CodeAlpaca dataset and subsequently subjected to structured pruning at 10%, 20%, 50%, and 75% compression levels. Recovery fine-tuning was applied after pruning to partially restore lost capabilities. Model performance was primarily evaluated using Pass@k on the HumanEval benchmark, while CodeBLEU and perplexity were employed as supplementary measures of code quality and language-modeling performance. Experimental results indicate that LLM compression exhibits strongly nonlinear behavior. Pass@k performance remained close to the fine-tuned baseline at low and moderate pruning levels (≤20%), suggesting that a substantial portion of model parameters can be removed with minimal impact on functional correctness. In contrast, performance deteriorated rapidly beyond a critical threshold near 50% pruning, where Pass@k values approached zero and perplexity increased significantly. CodeBLEU results further confirmed a substantial decline in code quality under aggressive compression. These findings reveal three operational regimes: a stable region (≤20% pruning), a transition region (20–50% pruning), and a collapse region (≥50% pruning), highlighting compression thresholds beyond which recovery fine-tuning becomes ineffective. These results demonstrate that compressed LLM behavior cannot be adequately understood through deterministic performance metrics alone and should instead be assessed using reliability-aware analysis that accounts for nonlinear degradation and stochastic variability. The framework provides practical guidance for selecting compression strategies that balance efficiency, robustness, and deployment feasibility.