<ul><li>Training data attribution (TDA) methods measure the impact of training data on a model's predictions.</li><li>Gradient-based attribution methods are theoretically grounded but computationally expensive, leading to impracticality for large-scale applications.</li><li>AirRep is introduced as a scalable, representation-based approach optimized for TDA by learning task-specific and model-aligned representations.</li><li>AirRep achieves comparable performance to gradient-based approaches but is significantly more efficient at inference time, showcasing robustness and generalization.</li></ul>

Enhancing Training Data Attribution with Representational Optimization

Discover more