7 papers · 1 filter
Analyzing and Improving Chain-of-Thought Monitorability Through Information Theory
Usman Anwar, Tim Bakker, Dana Kianfar +2
Chain-of-thought (CoT) monitors are LLM-based systems that analyze reasoning traces to detect when outputs may exhibit attributes of interest, such as test-hacking behavior during…
Understanding In-Context Learning of Linear Models in Transformers Through an Adversarial Lens
Usman Anwar, Johannes Von Oswald, Louis Kirsch +2
In this work, we make two contributions towards understanding of in-context learning of linear models by transformers. First, we investigate the adversarial robustness of in-contex…
Mitigating Goal Misgeneralization via Minimax Regret
Karim Abdel Sadek, Matthew Farrugia-Roberts, Usman Anwar +4
Safe generalization in reinforcement learning requires not only that a learned policy acts capably in new situations, but also that it uses its capabilities towards the pursuit of…
Interpreting Emergent Planning in Model-Free Reinforcement Learning
Thomas Bush, Stephen Chung, Usman Anwar +2
We present the first mechanistic evidence that model-free reinforcement learning agents can learn to plan. This is achieved by applying a methodology based on concept-based interpr…
Learning to Forget using Hypernetworks
Jose Miguel Lara Rangel, Stefan Schoepf, Jack Foster +2
Machine unlearning is gaining increasing attention as a way to remove adversarial data poisoning attacks from already trained models and to comply with privacy and AI regulations.…
Comparing Bottom-Up and Top-Down Steering Approaches on In-Context Learning Tasks
Madeline Brumley, Joe Kwon, David Krueger +2
A key objective of interpretability research on large language models (LLMs) is to develop methods for robustly steering models toward desired behaviors. To this end, two distinct…