Alexander Matt Turner

Output Supervision Can Obfuscate the Chain of Thought

Consistency Training Helps Stop Sycophancy and Jailbreaks

Mechanistically Eliciting Latent Behaviors in Language Models

Re-contextualization Mitigates Specification Gaming without Modifying the Specification

Distillation Robustifies Unlearning

Gradient Routing: Masking Gradients to Localize Computation in Neural Networks

Steering Language Models with Activation Engineering