CV — Adrià Garriga-Alonso

Profile

AI researcher and engineer, focused on alignment and mechanistic interpretability (2022–present), previously Bayesian ML (2017–2021). Senior author on subfield-defining interpretability work Automated Circuit Discovery (~400 citations), which sparked many follow-ups.

Combines deep research leadership with hands-on ML engineering: led teams of up to 6 researchers while building and maintaining critical infrastructure (8-80 GPU cluster, multi-node training) that accelerated research velocity at FAR AI. Apollo Research adopted the infrastructure as a reference design, and I consulted for Goodfire AI's infrastructure.

Seeking to understand how advanced AIs think, especially how they acquire, represent and follow goals.

Volunteer Research Mentorship

2024–2026

Program Mentor

ML Alignment & Theory Scholars

Advised 10 scholars on mechanistic interpretability and RL. Results: 3 NeurIPS papers, 5 workshop papers, 2 under review. Mentees progressed to positions at Anthropic, METR, and Mistral.

2025

Fellowship Mentor

Cambridge Boston Alignment Initiative

Advised 2 junior collaborators on 1 project: investigate whether LLM chain of thought answers are causal or rationalizations, using activation probes. Under review.

Professional Experience

Jun 2026–present

Contractor (part-time)

Theorem Labs

Guiding a team training AI to translate between two formal languages, from Rocq to Lean.

May 2026–present

Fellow

Mox

Researching coalitional agency, the predictive-coding view of intelligence, and agent foundations.

Feb–May 2026

Technical Co-founder

Dokimasia

Assisting humans in realizing their values when using the computer. Shield against unwanted and false information.

2023–2025

Research Scientist

FAR AI

Led two large projects on interpretability, directly managing 3 people and collaborating with 11.

Automated Circuit Discovery: FAR AI's most-cited work (~400 citations), which kicked off a new subfield in algorithmic circuit discovery.
Learned Planners. To study goal-directed behavior, we trained a model to play Sokoban using RL. Showed that the model learned to plan and internally represents future actions (ICLR Oral), and finally reverse-engineered the planning algorithm it learned.

Built and currently maintain FAR AI's GPU/LLM infrastructure (8-80 GPUs; >400B parameters), reducing costs by ~50%, enabling ≥20 researchers and collaborators to run large-scale experiments. Established core research project structure, engineering tools, and workflows now standard across the organization.

2022–2023

Member of Technical Staff

Redwood Research

Led correctness testing for an optimizing compiler and algebra system used by ~40 interns and 5 full-time researchers. Built fuzzing/property-based testing suite that uncovered critical bugs, ensuring reliability for experiments underpinning the Causal Scrubbing paper (co-author), and others. Mentored 8 interns across 4 projects, including Indirect Object Identification.

2021

Summer Research Fellow (open-source game theory)

Center on Long-Term Risk

What are the results of games when agents can read each other's source code? Proved that Nash-like equilibria are reached by probabilistic agents that attempt to prove each other's behavior, and that such behavior is described by modal logic.

2019

Research Intern

Microsoft Research Cambridge

With Dr. Sebastian Tschiatschek. Introduced a theoretically motivated algorithm to optimally choose and learn from partial observations of the teacher, in inverse reinforcement learning.

2014

CTO, cofounder

MonkingMe.com

Music streaming startup. Designed and implemented web application backend and cloud server infrastructure for the streaming service. Coordinated development of smartphone application.

Education

2017–2021

PhD in Machine Learning

Engineering Department, University of Cambridge, UK

Supervisor: Prof. Carl E. Rasmussen. Thesis: “Priors in finite and infinite Bayesian convolutional neural networks”. First to show that, at initialization, an infinitely wide non-MLP architecture converges to a Gaussian process. Proposed exact stochastic-gradient MCMC algorithm and improved priors for Bayesian neural networks.

2016–2017

MSc Computer Science (distinction)

University of Oxford, UK

Thesis supervisor: Prof. Mihaela van der Schaar. Thesis: “Probability density imputation of missing data with Gaussian mixture models”.

2012–2016

BSc Computer Science (1st in class of 67)

Pompeu Fabra University, Spain

Grade: 9.02/10, graduating class average 6.90/10. Thesis supervisor: Prof. Anders Jonsson. Thesis: “Solving Montezuma's Revenge with planning and reinforcement learning”.

Selected Publications

Towards Automated Circuit Discovery for Mechanistic Interpretability

A. Conmy, A. Mavor-Parker, A. Lynch, S. Heimersheim, A. Garriga-Alonso

NeurIPS 2023 Spotlight · ~400 citations

Deep Convolutional Networks as Shallow Gaussian Processes

A. Garriga-Alonso, L. Aitchison, C.E. Rasmussen

ICLR 2019 · ~330 citations

Open Problems in Mechanistic Interpretability

L. Sharkey, B. Chughtai, [...], A. Garriga-Alonso, et al.

arXiv 2025 · ~100 citations

Beyond the Imitation Game: Quantifying and Extrapolating the Capabilities of Language Models

A. Srivastava, A. Rastogi, A. Rao, [...], A. Garriga-Alonso, et al.

TMLR 2023 · ~2,000 citations

Causal Scrubbing: A Method for Rigorously Testing Interpretability Hypotheses

L. Chan, A. Garriga-Alonso, N. Goldowsky-Dill, R. Greenblatt, et al.

Alignment Forum 2022 · ~90 citations

Hypothesis Testing the Circuit Hypothesis in LLMs

C. Shi, N. Beltran Velez, A. Nazaret, C. Zheng, A. Garriga-Alonso, A. Jesson, et al.

NeurIPS 2024

Interpreting Emergent Planning in Model-Free Reinforcement Learning

T. Bush, S. Chung, U. Anwar, A. Garriga-Alonso, D. Krueger

ICLR 2025 Oral

Leadership

2019

Co-organizer, ICLR 2019 workshop: “Safe Machine Learning”

2017–2019

Founded and led Engineering AI Safety reading group

University of Cambridge

Objective: introduce ML students to beneficial AI techniques. 7–50 attendees per session.

Service to the Scientific Community

Reviewer

NeurIPS 2019 (top 5%), 2020, 2025. ICLR 2020, 2021, 2026. ICML 2020, 2021, 2023, 2025. JMLR. Workshops: ICML 2024 NextGen AI Safety, Mechanistic Interpretability (ICML 2024, NeurIPS 2025).

Mentor

New in ML workshop, 2019.

Selected Awards & Fellowships

2017

Malmö Collaborative AI Challenge: 1st & 3rd places, diff. categories

Microsoft Research

Won $20,000 in Azure credits and paid attendance to the AI Summer School.

2016

María de Maeztu Award for Reproducibility in Software

Pompeu Fabra University

Best computer science Bachelor's thesis in Spain meeting scientific reproducibility criteria.

2016–2017

Postgraduate fellowship (6.6% acceptance rate)

“la Caixa” Foundation

Full tuition and stipend. Awarded Spain-wide, on academic merit and positive impact of project.

Additional Activities

2014–2015

Competitive programming team member

Pompeu Fabra University

Conducted unofficial training sessions. Team set record in Pompeu Fabra University for number of problems solved, placing 24/52 in SWERC 2015.

Skills

Extensive experience with both PyTorch (most recently: interpretability projects) and JAX (RL for Sokoban). My freshest systems language is C++, but I can also write Rust.