A Geometric Account of Activation Steering through Angle–Norm Decomposition

·LessWrong··

This blog post provides an overview of our recent paper: A Geometric Account of Activation Steering through Angle–Norm Decomposition.TL;DR: We decompose linear activation steering into two distinct operations: one that changes the angle of the activation toward a concept direction, and one that changes its norm. Through controlled experiments, we analyze the role of each component. We find that concept information is indeed primarily encoded in the angular component of activations. However, the ...

Read full article →

Related Articles

Backdoor in Claude Code for VS Code extension on the Cursor Marketplace
_chall · Hacker News · 18h ago
Backdoors in official Claude Code extensions
dsr_ · Hacker News · 14h ago
Show HN: A working Nintendo in ~1GB of Qwen weights
dbreunig · Hacker News · 9h ago
We're releasing the weights of Mistral Large 4
kuberwastaken · Hacker News · 1d ago
Google Search is now less than half of Alphabet’s revenue
hubraumhugo · Hacker News · 1d ago