9 papers
Gradient Routing Localizes and Removes Unintended Behaviors in RL
Jake Ward, Shawn Hu, Aria Wong +4
We propose a mitigation to reward hacking in reinforcement learning. Our setup involves a misspecified reward function that gives rise to unintended behavior, and a monitor used to…
Modular Pretraining Enables Access Control
Ethan Roland, Murat Cubuktepe, Erick Martinez +8
AI developers face a dual-use dilemma. An AI capability that helps one user cure a disease can help another synthesize one. This dilemma could be resolved with access control, limi…
Beyond Data Filtering: Knowledge Localization for Capability Removal in LLMs
Igor Shilov, Alex Cloud, Aryo Pradipta Gema +5
Large Language Models increasingly possess capabilities that carry dual-use risks. While data filtering has emerged as a pretraining-time mitigation, it faces significant challenge…
Recontextualization Mitigates Specification Gaming without Modifying the Specification
Ariana Azarbal, Victor Gillioz, Vladimir Ivanov +6
Developers often struggle to specify correct training labels and rewards. Perhaps they don't need to. We propose recontextualization, which reduces how often language models "game"…
Natural Emergent Misalignment from Reward Hacking in Production RL
Monte MacDiarmid, Benjamin Wright, Jonathan Uesato +19
We show that when large language models learn to reward hack on production RL environments, this can result in egregious emergent misalignment. We start with a pretrained model, im…
Output Supervision Can Obfuscate the Chain of Thought
Jacob Drori, Luke Marks, Bryce Woodworth +2
OpenAI (2025) showed that training against a chain of thought (CoT) monitor can cause obfuscated CoTs, which contain bad behavior the monitor cannot detect. They proposed to keep C…
Subliminal Learning: Language models transmit behavioral traits via hidden signals in data
Alex Cloud, Minh Le, James Chua +5
We study subliminal learning, a surprising phenomenon where language models transmit behavioral traits via semantically unrelated data. In our main experiments, a "teacher" model w…
Distillation Robustifies Unlearning
Bruce W. Lee, Addie Foote, Alex Infanger +6
Current LLM unlearning methods are not robust. A few steps of finetuning can revert their effects. We begin by showing that this is true even for an idealized form of unlearning: t…
Gradient Routing: Masking Gradients to Localize Computation in Neural Networks
Alex Cloud, Jacob Goldman-Wetzler, Evžen Wybitul +2
Neural networks are trained primarily based on their inputs and outputs, without regard for their internal mechanisms. These neglected mechanisms determine properties that are crit…