Joan Velja

I work on adversarial elicitation and scalable oversight: how to reliably supervise systems smarter than the supervisor.

I'm a PhD student at the University of Oxford, supervised by Alessandro Abate and funded by the Alignment Project.

Selected research

Building methods for extracting knowledge from capable systems when the system may strategically resist supervision.

Alignment Project2026-present

When language models learn to coordinate through hidden signals, and what reliably prevents it.

IJCNLP-AACLOral - 2025

A safety-case framing for untrusted monitoring under realistic collusion assumptions.

Preprint2025

Preliminary Debate training improves proposal accuracy while exposing a critic strategy that learns to exploit the judge.

LessWrong25 min - 2026

What’s been in my ears lately.