Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2504.16574 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866912343130112000 |
|---|---|
| author | Chen, Lizhe Zhou, Binjia Ge, Yuyao Chen, Jiayi NI, Shiguang |
| author_facet | Chen, Lizhe Zhou, Binjia Ge, Yuyao Chen, Jiayi NI, Shiguang |
| contents | Large language models (LLMs) have achieved remarkable progress, demonstrating unprecedented capabilities across various natural language processing tasks. However, the high costs associated with such exceptional performance limit the widespread adoption of LLMs, highlighting the need for prompt compression. Existing prompt compression methods primarily rely on heuristic truncation or abstractive summarization techniques, which fundamentally overlook the intrinsic mechanisms of LLMs and lack a systematic evaluation of token importance for generation. In this work, we introduce Prompt Importance Sampling (PIS), a novel compression framework that dynamically compresses prompts by sampling important tokens based on the analysis of attention scores of hidden states. PIS employs a dual-level compression mechanism: 1) at the token level, we quantify saliency using LLM-native attention scores and implement adaptive compression through a lightweight 9-layer reinforcement learning (RL) network; 2) at the semantic level, we propose a Russian roulette sampling strategy for sentence-level importance sampling. Comprehensive evaluations across multiple domain benchmarks demonstrate that our method achieves state-of-the-art compression performance. Notably, our framework serendipitously enhances reasoning efficiency through optimized context structuring. This work advances prompt engineering by offering both theoretical grounding and practical efficiency in context management for LLMs. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2504_16574 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | PIS: Linking Importance Sampling and Attention Mechanisms for Efficient Prompt Compression Chen, Lizhe Zhou, Binjia Ge, Yuyao Chen, Jiayi NI, Shiguang Computation and Language Artificial Intelligence Large language models (LLMs) have achieved remarkable progress, demonstrating unprecedented capabilities across various natural language processing tasks. However, the high costs associated with such exceptional performance limit the widespread adoption of LLMs, highlighting the need for prompt compression. Existing prompt compression methods primarily rely on heuristic truncation or abstractive summarization techniques, which fundamentally overlook the intrinsic mechanisms of LLMs and lack a systematic evaluation of token importance for generation. In this work, we introduce Prompt Importance Sampling (PIS), a novel compression framework that dynamically compresses prompts by sampling important tokens based on the analysis of attention scores of hidden states. PIS employs a dual-level compression mechanism: 1) at the token level, we quantify saliency using LLM-native attention scores and implement adaptive compression through a lightweight 9-layer reinforcement learning (RL) network; 2) at the semantic level, we propose a Russian roulette sampling strategy for sentence-level importance sampling. Comprehensive evaluations across multiple domain benchmarks demonstrate that our method achieves state-of-the-art compression performance. Notably, our framework serendipitously enhances reasoning efficiency through optimized context structuring. This work advances prompt engineering by offering both theoretical grounding and practical efficiency in context management for LLMs. |
| title | PIS: Linking Importance Sampling and Attention Mechanisms for Efficient Prompt Compression |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2504.16574 |