Steven YangA collection of experiments and thoughts

I’m interested in what models learn about the world, and how those representations become actions. This is where I write up my experiments.

Changing a robot’s mind

Pick It Up: Does π0.5 Have an Editable Instruction Following Circuit?

·Interpretability / Robotics
● Object-centered● Random0.00.51.0481632641282565120.773Image positions replacedAction progress R

A late image-state edit redirects the robot to cream cheese. The position-dose curve shows why shrinking that edit proved difficult.

Identified a causal pathway for instruction control in π0.5, but no evidence for a selectively editable obedience circuit at the token position level.

Read the report →