Reproduce “Compress, Then Prompt: Improving Accuracy-Efficiency Trade-off of LLM Inference with Transferable Prompt” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhaozhuo Xu, Zirui Liu, Beidi Chen, Yuxin Tang, Jue Wang, Kaixiong Zhou, Xia Hu, Anshumali Shrivastava

A new paradigm, "Compress, Then Prompt," is introduced to enhance the performance of compressed Large Language Models (LLMs) through transferable soft prompts. This method allows aggressively compressed LLMs, such as 4-bit quantized and 50% pruned LLaMA-7B, to match or exceed the perplexity of their full, uncompressed counterparts while maintaining inference efficiency.

Key insights

  • LLMs demonstrate an intrinsic ability to adapt and self-correct their responses when explicitly prompted about their internal compressed state.
  • Learnable soft prompts offer a data-driven, precise method to consistently mitigate accuracy degradation in compressed LLMs.
  • The learned soft prompts exhibit broad transferability, working effectively across different datasets, compression levels/methods, and diverse downstream tasks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.