The Unreasonable Ineffectiveness of the Deeper Layers

TL;DR

This study investigates the importance of layers in large language models by pruning layers and observing minimal performance loss, suggesting that many deeper layers may be unnecessary for knowledge storage in common tasks.

Contribution

The paper introduces a layer pruning method combined with finetuning to assess the necessity of layers in LLMs, revealing surprising robustness to layer removal.

Findings

Up to half of the layers can be removed with minimal performance degradation.

Current pretraining may not fully utilize deeper layers for knowledge storage.

Shallow layers might be more critical than previously thought.

Abstract

How is knowledge stored in an LLM's weights? We study this via layer pruning: if removing a certain layer does not affect model performance in common question-answering benchmarks, then the weights in that layer are not necessary for storing the knowledge needed to answer those questions. To find these unnecessary parameters, we identify the optimal block of layers to prune by considering similarity across layers; then, to "heal" the damage, we perform a small amount of finetuning. Surprisingly, with this method we find minimal degradation of performance until after a large fraction (up to half) of the layers are removed for some common open-weight models. From a scientific perspective, the robustness of these LLMs to the deletion of layers implies either that current pretraining methods are not properly leveraging the parameters in the deeper layers of the network or that the shallow layers play a critical role in storing knowledge. For our study, we use parameter-efficient finetuning (PEFT) methods, specifically quantization and Low Rank Adapters (QLoRA), such that each of our experiments can be performed on a single 40GB A100 GPU.