Brian Christian

DPhil

DPhil (2023–2026)

My DPhil (2023–2026) was supervised by Christopher Summerfield (Experimental Psychology) and co-supervised by Jakob Foerster (Engineering Science) and Tsvetomira Dumbalska (Experimental Psychology).

My research interests are in the places where psychology and engineering meet: I’m interested in computational models of human cognition, in the structure and representation of human rewards and goals, and in reward models and reinforcement learning from human feedback (RLHF) as promising, but incomplete, tools for operationalizing notions of human norms, preferences, and values.

DPhil Thesis: 'What humans want: Models of preference, choice, and reward in minds and machines'.

Cookies on this website

Contact information

Collaborators

Tsvetomira Dumbalska

Matan Mazor

Jessica Thompson

Kai Sandbrink

Colleges

Brian Christian

Key publications

Reward Model Interpretability via Optimal and Pessimal Tokens

The Alignment Problem Machine Learning and Human Values

Revealing priors on category structures through iterated learning

Recent publications

Resolving Feynman's restaurant problem reveals optimal solutions and human strategies.

Reward models inherit value biases from pretraining

Reward Model Interpretability via Optimal and Pessimal Tokens

Using adaptive intrinsic motivation in RL to model learning across development

Computational Frameworks for Human Care