Understand the risk.
Investigate vulnerabilities that remain hidden behind safe-looking answers. Develop representation-level evaluations for open-weight models.
Evaluation & diagnosisAn emerging AI safety research community
Look inside.
Build safer AI.
A model’s answers tell only part of the story. We study what happens beneath the surface—and how that understanding can make AI safer.
Started by Enyi (Olivia) JiangA small community, working on shared research questions.
BEYOND THE OUTPUTFIG. 01
Find the hidden riskCONCEPTUAL VIEW
01 / Research directions
Safety, from the inside out.
We connect evaluation, alignment, and intervention through the internal representations of AI models.
Shape the learning.
Explore how feedback from internal representations can improve alignment and make safety more robust under changing conditions.
Representation-based alignmentIntervene with care.
Build selective monitoring and steering methods that respond to emerging risks while preserving useful capabilities and task progress.
Monitoring & controlOne model. Three research lenses.
Evaluation asks where apparently safe behavior breaks down.
02 / Selected work
A foundation to build on.
Research by our founder and collaborators that informs this initiative’s agenda. These works predate the organization.
When Behavioral Safety Evaluation Fails
A Representation-Level Perspective
Studying the gap between passing behavioral audits and resisting internal interventions.
Latent Adversarial Regularization for Offline Preference Optimization
GANPO explores latent-space feedback for more robust preference learning.
CLEAR: Continuous Latent Adapter Routing
Utility-Preserving LLM Safety Alignment
Adjusting safety interventions using a model’s hidden states.
Come say hello.
Have an idea?
Let’s talk.
Tell us what you’re working on, what you’d like to learn, or how you’d like to get involved.
Researchers · Students · Organizers · Research partners