E. Lubana

This work finds that simple difference of means vectors coherently represent reward hacking in Kimi K3, GLM 5.2, and Qwen 3.8 Max, and provides evidence that simple, white-box methods can be used to scalably study and monitor reward hacking behaviors in frontier open source models.

Uncertainty in LLM reasoning is largely a sampling artifact, allowing for more efficient analyses that cut costs without sacrificing accuracy.

Larger models learn more not just because of increased capacity, but because they experience less interference during training, allowing them to retain rare and complex tasks that smaller models forget.

Sparse autoencoders, despite their popularity for extracting interpretable features, often fail to capture the underlying manifold structure of concepts, instead fragmenting them across multiple, diluted features.