This paper addresses the issue of accountability in modern AI systems, which undergo multi-stage development processes (pretraining, fine-tuning, and adaptation/alignment). We address the "attribution problem"—determining which development stage is responsible for the success or failure of a deployed model, and propose a general framework to answer the semi-empirical question of how the model's behavior would have changed had a specific stage not been updated. Within this framework, we introduce an estimator that efficiently quantifies the effects of each stage without requiring model retraining, taking into account key aspects of model optimization dynamics, such as learning rate schedules, momentum, and weight decay. We demonstrate that this method successfully quantifies accountability for model behavior at each stage and identifies and removes learned spurious correlations in multi-stage development tasks such as image classification and text toxicity detection. In conclusion, this paper provides a practical tool for model analysis and represents an important step toward more responsible AI development.
