William Lugoloobi

I train and build reliable AI agents.

I'm William Lugoloobi, a third-year PhD student at the Oxford Internet Institute, advised by Joss Wright and Chris Russell. My research spans model self-knowledge, post-training, and harness optimisation for long-horizon agents. I'm currently a Research Scientist Intern at Thomson Reuters.

Selected publications ↓Google Scholar ↗GitHub ↗
William LugoloobiOxford · London

Selected publications

All on Scholar ↗
Known By Their Actions: Fingerprinting LLM Browser Agents via UI Traces

Websites can silently fingerprint which LLM is powering a browser agent just by observing its actions with an embedded tracker.

arXiv:2605.14786 · Computer-use agents · Security
Paper ↗
LLMs Encode Their Failures: Predicting Success from Pre-Generation Activations

Language models’ internal representations contain signal about whether they’ll succeed on a task before generation begins. We use it to route inference efficiently while maintaining performance.

COLM 2026 · Mechanistic interpretability · Inference routing
Paper ↗
LLMs Encode How Difficult Problems Are

LLMs linearly represent the difficulty of math and coding problems in their activation space, with implications for adaptive inference, curriculum learning, and model confidence.

arXiv:2510.18147 · Representation learning · Probing
Paper ↗

Selected projects

Live

pidge.watch

Tell us what Oxford role you want. Know when it appears.

Oxford jobs are scattered across dozens of college sites and can appear without warning. Tell Pidge the role, college, or perk you want—it watches for matches and alerts you the moment one appears.

Visit Pidge
Open source

TrajectoryKit

Local-first agent observability.

Run agents on your own hardware and inspect every turn, tool call, reasoning chain, and sub-agent trace.

#6 on DeepResearch BenchGitHub

News

Recent work & updates

LLMs Encode Their Failures accepted to COLM Our work shows that models encode whether they will succeed before generation begins, enabling more efficient inference routing.

Joined Thomson Reuters as a Research Scientist Intern Working on post-training and harness optimisation for agents.

Launched Pidge Search Oxford college roles in one place and follow the opportunities, perks, and deadlines you care about.

New paper on browser-agent fingerprinting We show that websites can identify the model behind an LLM browser agent from its UI traces.

Interested in reliable agents, post-training, or collaborations?

william.lugoloobi [at] oii.ox.ac.ukLinkedIn ↗CV ↗