What Makes Two Language Models Think Alike?
Researchers have developed a new method to understand why language models produce similar or different outputs. By mapping neural activity to linguistic features, they can quantify the drivers of similarity between models. Their analysis of 43 models across various families revealed that release date and model family, rather than scale or architecture class, most strongly influence model-level similarity. AI
IMPACT Provides a new analytical tool to understand model behavior and potential biases, aiding in model selection and development.