Picture for Max Nadeau

Max Nadeau

Circuit Breaking: Removing Model Behaviors with Targeted Ablation

Add code
Sep 12, 2023
Viaarxiv icon

Benchmarks for Detecting Measurement Tampering

Add code
Sep 07, 2023
Viaarxiv icon

Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback

Add code
Jul 27, 2023
Viaarxiv icon

Discovering Variable Binding Circuitry with Desiderata

Add code
Jul 07, 2023
Viaarxiv icon

One Thing to Fool them All: Generating Interpretable, Universal, and Physically-Realizable Adversarial Features

Add code
Oct 11, 2021
Figure 1 for One Thing to Fool them All: Generating Interpretable, Universal, and Physically-Realizable Adversarial Features
Figure 2 for One Thing to Fool them All: Generating Interpretable, Universal, and Physically-Realizable Adversarial Features
Figure 3 for One Thing to Fool them All: Generating Interpretable, Universal, and Physically-Realizable Adversarial Features
Figure 4 for One Thing to Fool them All: Generating Interpretable, Universal, and Physically-Realizable Adversarial Features
Viaarxiv icon