research
lab notebook. running notes and results from my interpretability research.
-
Tracing a Fine-Tuned SmolVLA from Pixels to Action
-
Consider Performance for Better Science
thoughts on emergent behavior in science.
-
A Complete Mechanistic Account of Single-Joint Control in SmolVLA
A research note in mechanistic interpretability.
-
Cross-Language Transfer of Harm Probes in Vision-Language-Action Models
A research note on multilingual harm detection.
-
English–Chinese Code-Switched Instructions in a MiniVLA Policy
A paired study of verb language, noun language, and sentence frame in robot instructions.
-
What a Size Ladder Can Identify
why the scaling exponents optimizer benchmarks fit measure the experimental design, not the optimizer
-
Reconstructing ROME
how we converged on ROME, from a new direction
-
Playing with VPD
what I discovered when playing with VPD
-
Sensor Context Encoder Challenge
Can a window of inertial-sensor data be handed to a frozen language model as one continuous embedding — never rendered as text — and read back out as an activity label?
-
Engineering Notebook Entry
I reinvented ROME & pivot
-
Engineering Notebook Entry
working with a toy model
-
Engineering Notebook Entry
exploring new techniques
-
TartanIMU
from my robotics portfolio, linked here for completeness.