Large language models (LLMs) provide excellent performance across a wide range of natural language tasks, but their practical deployment in real-world systems is often limited by significant inference costs and latency under strict resource constraint...
Large language models (LLMs) provide excellent performance across a wide range of natural language tasks, but their practical deployment in real-world systems is often limited by significant inference costs and latency under strict resource constraints. While block pruning has emerged as an effective strategy to reduce computation and preserve the structural coherence of transformer architectures, existing methods typically rely on representation similarity or costly sensitivity analyses, which only weakly reflect task-aware model behavior and are difficult to scale in practice. In this thesis, Task-aware Block Pruning (TaBP) is proposed as a novel block pruning framework that quantifies block-level uncertainty by attaching lightweight LM heads to each block and measuring entropy-based statistics of their early-exited output distributions on a task-specific calibration dataset, in order to identify prunable blocks. Extensive experiments on multiple LLM backbones and downstream benchmarks validate the effectiveness of the proposed method, demonstrating substantial efficiency gains without compromising task performance, while avoiding computationally expensive sensitivity analyses.