Enterprise AI Analysis

QASPC: Question-Aware Sentence-Level Prompt Compression

Large language models (LLMs) rely heavily on very long prompts to achieve strong performance in complex reasoning and understanding tasks, but such prompts introduce substantial computational and financial overhead. Prompt compression has therefore become an essential direction for improving inference efficiency, yet existing approaches often require additional training or operate at the token level, leading to fragmented outputs and degraded task performance under high compression ratios. To address these limitations, we propose Question-Aware Sentence-Level Prompt Compression (QASPC), a fully training-free, two-stage framework that integrates sentence-level semantic filtering with clause-level perplexity and question relevance analysis. QASPC first identifies a maximally informative and coherent subset of sentences, and then performs fine-grained clause selection based on sliding-window perplexity and question-aware scoring. Experiments on the LongBench benchmark indicate that QASPC achieves strong generalization across long-text understanding tasks and consistently outperforms existing compression baselines in both single-document QA and summarization. These results highlight the effectiveness of semantic-level compression for reducing input length while preserving essential contextual information.