SelfCP: Compressing over-limit prompt via the frozen large language model itself

Abstract

Long prompt leads to huge hardware costs when using transformer-based Large Language Models (LLMs). Unfortunately, many tasks, such as summarization, inevitably introduce long documents, and the wide application of in-context learning easily makes the prompt length explode. This paper proposes a Self-Compressor (SelfCP), which adopts the target LLM itself to compress over-limit prompts into dense vectors on top of a sequence of learnable embeddings (memory tags) while keeping the allowed prompts unmodified. Dense vectors are then projected into memory tokens via a learnable connector, allowing the same LLM to understand them. The connector and the memory tag are supervised-tuned under the language modeling objective of the LLM on relatively long texts selected from publicly accessed datasets involving an instruction dataset to make SelfCP respond to various prompts, while the target LLM keeps frozen during training. We build the lightweight SelfCP upon 2 different backbones with merely 17M learnable parameters originating from the connector and a learnable embedding. Evaluation on both English and Chinese benchmarks demonstrate that SelfCP effectively substitutes 12× over-limit prompts with memory tokens to reduce memory costs and booster inference throughputs, yet improving response quality. The outstanding performance brings an efficient solution for LLMs to tackle long prompts without training LLMs from scratch.