Is Bigger and Deeper Always Better? Probing LLaMA Across Scales and Layers

TL;DR

This study probes LLaMA models across different sizes and layers using targeted tasks, revealing that larger models improve reasoning and reduce hallucinations only beyond certain sizes, while top layers hold most knowledge.

Contribution

It introduces a novel probing approach to analyze LLaMA's internal representations across scales and layers, revealing nuanced insights into model capabilities.

Findings

Larger models enhance reasoning and reduce hallucinations beyond certain sizes.

Lower layers lack arithmetic and factual knowledge, focusing on logical and multilingual abilities.

Top layers contain most computational power and real-world knowledge.

Abstract

This paper presents an in-depth analysis of Large Language Models (LLMs), focusing on LLaMA, a prominent open-source foundational model in natural language processing. Instead of assessing LLaMA through its generative output, we design multiple-choice tasks to probe its intrinsic understanding in high-order tasks such as reasoning and computation. We examine the model horizontally, comparing different sizes, and vertically, assessing different layers. We unveil several key and uncommon findings based on the designed probing tasks: (1) Horizontally, enlarging model sizes almost could not automatically impart additional knowledge or computational prowess. Instead, it can enhance reasoning abilities, especially in math problem solving, and helps reduce hallucinations, but only beyond certain size thresholds; (2) In vertical analysis, the lower layers of LLaMA lack substantial arithmetic and factual knowledge, showcasing logical thinking, multilingual and recognitive abilities, with top layers housing most computational power and real-world knowledge.