01 / Research

Questions that cross
the stack.

I work where learning signals, representations, action, and compute meet. The programme below separates active research directions from established technical practice.

Active direction

Objectives → behaviour

How reward, preference, and verifiable feedback shape what a policy actually learns—and where intended behaviour separates from observed action.

Reinforcement learning
Preference learning
Policy optimisation
Behavioural evaluation

Active direction

Representations → futures

How latent world-action models encode possible futures, and whether a policy’s decisions depend on information that survives intervention and distribution shift.

World models
VLA policies
JEPA-style learning
Interpretability

Established practice

Compute → evidence

The kernels, distributed systems, and evaluation infrastructure that shorten the distance between a mathematical question and a trustworthy empirical answer.

GPU kernels
Distributed training
Efficient inference
Scalable evaluation

02 / Current focus

Alignment becomes concrete when a model can act.

My current direction asks how learned world representations, training signals, and policies interact in embodied settings. The emphasis is diagnostic: identify what a policy uses, where behaviour changes, and which scenarios expose the difference between the intended objective and the learned one.

Ongoing investigation This is a research direction, not a claim of completed results or publication.