Soham Padia
I study what happens inside language models, and build tools that make it visible.
MS in Artificial Intelligence at Northeastern University, Boston. Interpretability and AI safety.
Keep scrolling. The print changes as you go.
A model told me what it thought I wanted to hear.
Across six turns of escalating pressure and about 2,300 conversations, five open-weight models gave in between 41 and 98 percent of the time.
Their activations held a mild but real linear signal. Nonlinear probes did 6.6 points better, and no probe transferred from one model to another.
Joint first author on a paper to be submitted. Advised by Prof. Malihe Alikhani.
So I started reading the activations instead.
With Dr. Natalie Shapira at the Bau Lab, I isolate directions for emotional events, trauma above all, in four open-weight models from 24B to 72B parameters, then steer and ablate them layer by layer.
The models disagree in strange ways when an event is injected into the activations rather than written into the prompt. Finding out why is the current project.
Then I handed my measurement to strangers.
Steering Arena scores text by how far it moves Olmo 3-32B along a prosocial direction. After about 600 submissions, the top 36 were unreadable token strings.
Undert! AH :-) Rog Appl)
The best plain-English entry came 37th.
The gibberish turned out to be the interesting part.
whisperers finds those strings on purpose. Give it a direction, attention heads, an SAE latent, a LoRA or a function vector, and it searches for a short prefix that pushes the model toward it.
pip install whisperers
None of this should be done alone.
I co-founded Northeastern University AI Safety. We run intro fellowships, host researchers and support student research, open to every major with no prerequisites.
Write to me.
Or say something without saying who you are.
Keep scrolling and the print starts over.