Physics of Language Models: Part 3.2, Knowledge Manipulation

TL;DR

This paper examines the ability of language models to manipulate stored knowledge across tasks like retrieval, classification, comparison, and inverse search, revealing significant limitations even in large models like GPT-4.

Contribution

It provides a controlled experiment demonstrating that language models inherently struggle with knowledge manipulation tasks, highlighting fundamental weaknesses despite extensive training.

Findings

Models excel in knowledge retrieval but struggle with classification and comparison.

Performance in inverse search is nearly zero, regardless of prompts.

Weaknesses are inherent, not due to model size or training data.

Abstract

Language models can store vast factual knowledge, yet their ability to flexibly use this knowledge for downstream tasks (e.g., via instruction finetuning) remains questionable. This paper investigates four fundamental knowledge manipulation tasks: retrieval (e.g., "What is person A's attribute X?"), classification (e.g., "Is A's attribute X even or odd?"), comparison (e.g., "Is A greater than B in attribute X?"), and inverse search (e.g., "Which person's attribute X equals T?"). We show that language models excel in knowledge retrieval but struggle even in the simplest classification or comparison tasks unless Chain of Thoughts (CoTs) are employed during both training and inference. Moreover, their performance in inverse knowledge search is virtually 0%, regardless of the prompts. Our primary contribution is a controlled, synthetic experiment that confirms these weaknesses are inherent to language models: they cannot efficiently manipulate knowledge from pre-training data, even when such knowledge is perfectly stored in the models, despite adequate training and sufficient model size. Our findings also apply to modern pretrained language models such as GPT-4, thus giving rise to many Turing tests to distinguish Humans from contemporary AIs.