English

Do Compressed LLMs Forget Knowledge? An Experimental Study with Practical Implications

Computation and Language 2024-02-19 v3 Artificial Intelligence

Abstract

Compressing Large Language Models (LLMs) often leads to reduced performance, especially for knowledge-intensive tasks. In this work, we dive into how compression damages LLMs' inherent knowledge and the possible remedies. We start by proposing two conjectures on the nature of the damage: one is certain knowledge being forgotten (or erased) after LLM compression, hence necessitating the compressed model to (re)learn from data with additional parameters; the other presumes that knowledge is internally displaced and hence one requires merely "inference re-direction" with input-side augmentation such as prompting, to recover the knowledge-related performance. Extensive experiments are then designed to (in)validate the two conjectures. We observe the promise of prompting in comparison to model tuning; we further unlock prompting's potential by introducing a variant called Inference-time Dynamic Prompting (IDP), that can effectively increase prompt diversity without incurring any inference overhead. Our experiments consistently suggest that compared to the classical re-training alternatives such as LoRA, prompting with IDP leads to better or comparable post-compression performance recovery, while saving the extra parameter size by 21x and reducing inference latency by 60%. Our experiments hence strongly endorse the conjecture of "knowledge displaced" over "knowledge forgotten", and shed light on a new efficient mechanism to restore compressed LLM performance. We additionally visualize and analyze the different attention and activation patterns between prompted and re-trained models, demonstrating they achieve performance recovery in two different regimes.

Keywords

Cite

@article{arxiv.2310.00867,
  title  = {Do Compressed LLMs Forget Knowledge? An Experimental Study with Practical Implications},
  author = {Duc N. M Hoang and Minsik Cho and Thomas Merth and Mohammad Rastegari and Zhangyang Wang},
  journal= {arXiv preprint arXiv:2310.00867},
  year   = {2024}
}