RADAR: Relative Angular Divergence Across Representations
Researchers have developed RADAR, a new metric designed to estimate the transferability of foundation models across different domains. This method analyzes the geometric evolution of representations within a model's layers to predict how well it will perform on new, unseen data. RADAR has shown competitive performance against existing metrics in both text and image classification tasks, particularly when domain shifts are clear. AI
IMPACT Provides a new tool for evaluating how well foundation models will adapt to new data, potentially guiding model selection and fine-tuning efforts.