Does Instruction Tuning Make LLMs More Consistent?

TL;DR

This paper investigates how instruction tuning affects the consistency of large language models, showing that it generally improves stability in predictions and representations across various tasks.

Contribution

It provides a comparative analysis demonstrating that instruction tuning enhances model consistency, supported by mechanistic insights into factual recall.

Findings

Instruction tuning increases model consistency in predictions.

Models show improved stability in zero-shot and downstream tasks.

Mechanistic analysis explains the basis for increased consistency.

Abstract

The purpose of instruction tuning is enabling zero-shot performance, but instruction tuning has also been shown to improve chain-of-thought reasoning and value alignment (Si et al., 2023). Here we consider the impact on $\textit{consistency}$, i.e., the sensitivity of language models to small perturbations in the input. We compare 10 instruction-tuned LLaMA models to the original LLaMA-7b model and show that almost across-the-board they become more consistent, both in terms of their representations and their predictions in zero-shot and downstream tasks. We explain these improvements through mechanistic analyses of factual recall.