David Krueger

We highlight the susceptibility of existing unlearning methods to relearning attacks and analyze the characteristics of robust methods by leveraging the weight-space perspective.

We train probes on activations to classify high- vs low-stakes scenarios, find they outperform medium-sized fine-tuned LLMs, and consider applications to monitoring.

This paper introduces distributional training data attribution, a data attribution framework that accounts for stochasticity in deep learning training, enabling a mathematical justification for why influence functions work in this setting.