Ben Bethell

Machine Learning · Interpretability · Evaluations

Hi! I'm Ben Bethell, a machine learning researcher and engineer.

I'm interested in understanding and evaluating machine learning systems, with a focus on safety and real-world use.

I recently completed an MSc in Data Science at Birkbeck, University of London. I live in London.

GitHub LinkedIn

Selected projects

Ladder chart for LLaMA-3-8B: edits are recalled at 95% but captured headroom falls with each step away from the edit, to about 0% by multi-hop and logical probes

Model Editing: Capacities and Limits

Two-Tower Document Retrieval

Research

Model editing: capacities and limits

MSc thesis · Birkbeck · 2026

Do model-editing methods change what a language model knows, or only what it recalls? I implemented six editing methods (ROME, MEMIT, AlphaEdit, AnyEdit, GRACE, REMEDI) under one evaluation harness on GPT-2-XL and LLaMA-3-8B, and built GENIE, a benchmark that tests whether an edit is used rather than just recalled. Every method installs the edit at 92.6–100% on CounterFact; the best score two rungs of abstraction later is 12.6%.

Code

Writing

First post coming soon.