Our work on ICP for Data Attribution was accepted to NAACL 2025. We showed that simple probing of LLMs may serve as a practical proxy for gradient-based data attribution, enabling efficient identification of influential training samples. Check out our paper and code.
Our work on ICP for Data Attribution was accepted to NAACL 2025. We showed that simple probing of LLMs may serve as a practical proxy for gradient-based data attribution, enabling efficient identification of influential training samples. Check out our 
paper
code