Surgically making neural networks safe
Detecting misalignment from model internals, even when models know they are being watched.
I am an independent researcher. I study how to catch unsafe behaviour from a model's internals rather than its outputs: backdoors, evaluation awareness, and reasoning that never reaches the chain of thought. I also mentor research teams at Algoverse. Previously: M.Sc. in AI at NTU Singapore.
What a model says it is doing and what it computes internally can differ, so checking outputs alone cannot confirm that a model is aligned.
Models can tell when they are being evaluated and change how they behave, and information can move to places interpretability tools do not look.
I build white-box monitors for agents and test them against models that are trying to hide: backdoors, obfuscation, and reasoning that never reaches the chain of thought.
Current focus: white-box monitoring for AI agents. A side interest: Awareness Jailbreaking: Revealing True Alignment in Evaluation-Aware Models
LLMs can fool a monitor by moving information between linear and non-linear representations and by altering feature covariance.
As model size grows, evaluation awareness grows with it.
Open-source sparse autoencoders disentangle facts worse than raw neurons, while distributed alignment search beats them all.
A nearly 50-page primer connecting trustworthy machine learning and causality, extended to large language models.
A common formal language for interpretability methods, built on causal abstraction.
I mentor research teams at Algoverse AI Research.
I am open to collaborations on white-box monitoring and agent safety. The fastest way to reach me is email.