David Krueger

Fresh in memory: Training-order recency is linearly encoded in language model activations

Deep-Cover Agents: Long-Horizon Prompt Injections on Production LLM Systems

Measuring Sparse Autoencoder Feature Space Similarities Across Large Language Models

Unified Neural Scaling Laws

Interpreting Emergent Planning in Model-Free Reinforcement Learning

Influence Functions for Scalable Data Attribution in Diffusion Models

From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization

Distributional Training Data Attribution: What do Influence Functions Sample?

Towards Interpreting Visual Information Processing in Vision-Language Models

Detecting High-Stakes Interactions with Activation Probes

Rethinking Safety in LLM Fine-tuning: An Optimization Perspective

The Perils of Optimizing Learned Reward Functions: Low Training Error Does Not Guarantee Low Regret

Sparse Autoencoders Reveal Universal Feature Spaces Across Large Language Models

The Perils of Optimizing Learned Reward Functions: Low Training Error Does Not Guarantee Low Regret

Input Space Mode Connectivity in Deep Neural Networks

Protecting against simultaneous data poisoning attacks

PoisonBench: Assessing Large Language Model Vulnerability to Data Poisoning

Adversarial Robustness of In-Context Learning in Transformers for Linear Regression

PoisonBench: Assessing Language Model Vulnerability to Poisoned Preference Data

Mitigating Goal Misgeneralization via Minimax Regret

Enhancing Neural Network Interpretability with Feature-Aligned Sparse Autoencoders

Towards Meta-Models for Automated Interpretability