William Lugoloobi

Hi, I'm
William

I'm a 2nd-year PhD student at the Oxford Internet Institute, advised by Joss Wright and Chris Russell. My work explores how language models internally represent their own limitations, and how to build AI agents that stay safe and on task over long horizons.

AI SafetyAgentic SystemsInterpretability
William Lugoloobi

Research

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 show this can be extracted to route inference efficiently, reducing costs while maintaining performance.

(under review, 2026)  ·  mechanistic interpretability  ·  inference routing
Feb 2026
→
LLMs Encode How Difficult Problems Are

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

arXiv:2510.18147  ·  representation learning  ·  probing
Oct 2025
→

Projects

TrajectoryKit

A local-first agentic framework for running AI agents on your own hardware (BYOK). Built on the idea that full observability is prerequisite to safety: every turn, reasoning chain, tool call, and sub-agent is captured in a structured trace and rendered as a readable HTML report.

dispatch_trace
25e4faa4 · Qwen/Qwen3-8B · local
176sDuration
40Turns
10Sub-agents
141KTokens
root · turn 1Decomposing task → spawning 3 sub-agents in parallel for card stats lookup
sub-agent · depth 1search_web("Blue-Eyes White Dragon stats") → {atk: 3000, def: 2500, level: 8}
tool callexecute_code(matplotlib) → bar chart with ★ annotation on strongest monster
sub-agent · depth 2search_web("spell card ATK boost") → Legend of Blue-Eyes White Dragon (+300 ATK)
root · turn 8Synthesizing enhanced stats → rendering final comparison
100% localRuns on your own GPU via vLLM. No API keys, no cloud dependency.
recursive sub-agentsSpawn child agents to decompose complex tasks independently, each with its own nested trace, up to 3 levels deep.
sandboxed executionRun Python and 40+ languages safely via Apptainer, with file I/O and image retrieval.
HTML trace viewerSelf-contained trace reports with collapsible reasoning chains, inline images, and full token accounting.
GitHub →PyPIpip install trajectorykit

Contact

Contact me by email at X@Y where X=william.lugoloobi and Y=oii.ox.ac.uk