Sparse Autoencoders Can Learn Graded Latents for Relational Composition
Theo Farrell, Patrick Leask, Noura Al Moubayed
Mechanistic Interpretability Workshop at ICML 2026
View on OpenReview
AI Safety Research and Field-building
Theo Farrell
MSci Natural Sciences (Computer Science & Philosophy) at Durham University. Founder and co-organiser of Durham AI Safety Initiative. I’ve got experience in foundational mechanistic interpretability research, but I’m increasingly interested and concerned about the risks from open weight models (as outlined in this paper). Starting summer 2026 I will be in London for the LASR Labs fellowship. Outside of AI Safety I play bass guitar 🎸 and dabble with piano and viola 🎹🎻!
Research
Published papers I played a key role in creating:
Order by Scale: Relative-Magnitude Relational Composition in Attention-Only Transformers
Theo Farrell, Patrick Leask, Noura Al Moubayed
39th Conference on Neural Information Processing Systems (NeurIPS 2025) Workshop: ResponsibleFM
View on OpenReviewOther published papers I’ve helped with:
Challenges of Evaluating LLM Safety for User Welfare
Manon Kempermann, Sai Suresh Macharla Vasu, Mahalakshmi Raveenthiran, Theo Farrell, Ingmar Weber
Second Conference of the International Association for Safe and Ethical Artificial Intelligence (IASEAI'26)
View preprint on arXivReviewing
- ICML 2026 Mechanistic Interpretability Workshop
- NeurIPS 2025 Mechanistic Interpretability Workshop
- NeurIPS 2025 ResponsibleFM Workshop
Talks
See All Talks →AI Safety North East
5 May 2026 • Centre for AI Safety, Newcastle University
How reliable are current interpretability methods? Recent work on CoT monitoring and SAEs (co-presented with Toby Pullan)
Field-building
Durham AI Safety Initiative
Lead Organiser · October 2023 — OngoingFounding and scaling an AI Safety university group
I founded DAISI in my second year of university after being inspired by a three-person reading group at the university’s Effective Altruism society. Supported by the Pathfinder fellowship, the group helps to funnel top-university talent into AI Safety.
- Organised recruitment and increased our weekly attendance from 5 to 20 participants.
- Facilitated BlueDot discussion groups to upskill members in AI alignment and policy knowledge.
- Led technical upskilling workshops in deep learning and NLP based on ARENA.
- Secured grant funding for group organising.
- Pathfinder Fellow since Dec 2024 - received training and mentorship for AI Safety organising after a competitive application process.
Projects
Role
Category
Languages
Tools
Focus
Sort by
Bipedal Walking with RL
Reinforcement learning implementation for training bipedal agents to walk
lead coursework
AI Safety
Challenges of Evaluating LLM Safety for User Welfare
IASEAI 2026
Challenges of Evaluating LLM Safety for User Welfare — Code & Dataset
contributor research
Cryptography Coursework
Algorithm and protocol implementation and analysis
lead coursework
Data Cleaning and Analysis
Data preprocessing, cleaning, and statistical analysis
lead coursework
Data Compression Coursework
Implementation of data compression algorithms including Huffman coding, LZW, and run-length encoding
lead coursework
AI Safety
Durham AI Safety Initiative Website
Organisation website with information on DAISI initiatives and research
lead side-project
EPQ Calculator
First major programming project — a calculator application for learning fundamentals
lead coursework
AI Safety
Ghost Marks in the Machine
d/acc Hackathon @ Apart Research
Research project investigating watermarking LLM code output
lead side-project
Hands2Text
Hackathon project converting sign language to text using computer vision and machine learning (DurHack 2023)
lead side-project
Human Features for Style Transfer
Deep learning pipeline enhancing human character visuals in game videos using movie reference footage. Integrates YOLO detection, GCN pose classification, and patch-level CUT.
lead coursework
Image Processing
Computer vision and image processing techniques
lead coursework
KNN vs Logistic Regression
Comparison of K-Nearest Neighbours and Logistic Regression classifiers on various datasets
lead coursework
M
Metaheuristics and TSP
Implementation and comparison of metaheuristic algorithms (genetic algorithm and ant colony optimisation) for the Travelling Salesman Problem
lead coursework
N
Natural Computing Algorithms
Exploration of bio-inspired and nature-based computational algorithms including cuckoo search and negative selection
lead coursework
AI Safety
Order by Scale
ResponsibleFM @ NeurIPS 2025
Implementation of research on relative-magnitude relational composition in attention-only transformers
lead research
Programmable Matter Lab
Interactive browser simulators for bio-inspired computing models, featuring active membranes (P-systems) and connectivity-preserving robot swarm shape transformations.
lead side-project
Stance Classification
NLP project implementing transformer-based models for detecting argumentative stances in text
lead coursework
Let's chat!
Pick a time that works for you and I'll see you there.