Identifying Linear Relational Concepts in Large Language Models

2023/11/15 by D.I. Chanin, Chanin, David, Anthony Hunter +3 · 3 citations
Computer Science · Social Sciences · #Topic Modeling #Natural Language Processing Techniques #Computational and Text Analysis Methods

paper · pdf · doi:10.48550/arxiv.2311.08968

Abstract

Transformer language models (LMs) have been shown to represent concepts as directions in the latent space of hidden activations. However, for any human-interpretable concept, how can we find its direction in the latent space? We present a technique called linear relational concepts (LRC) for finding concept directions corresponding to human-interpretable concepts by first modeling the relation between subject and object as a linear relational embedding (LRE). We find that inverting the LRE and using earlier object layers results in a powerful technique for finding concept directions that outperforms standard black-box probing classifiers. We evaluate LRCs on their performance as concept classifiers as well as their ability to causally change model output.

Cited by

• Task Matrices: Linear Maps for Cross-Model Finetuning Transfer
• Behavior and Representation in Large Language Models for Combinatorial Optimization: From Feature Extraction to Algorithm Selection
• The Structure of Relation Decoding Linear Operators in Large Language Models

Related

• Linearity of Relation Decoding in Transformer Language Models
• Beyond Single Concept Vector: Modeling Concept Subspace in LLMs with Gaussian Distribution
• Cross-model Transferability among Large Language Models on the Platonic Representations of Concepts
• Linear Relational Decoding of Morphology in Language Models
• The Linear Representation Hypothesis and the Geometry of Large Language Models