Adrian Goldwaser
About Me
I’m an ML researcher focused on understanding what neural networks are actually doing: how they learn, why they generalise, how they use context, and how that understanding can make them more interpretable, steerable and safe.
Most recently, I showed that in-context learning in all widely used transformer architectures is exactly equivalent to an implicit weight update (ICML 2026, oral), and with colleagues at Google applied this to steering how LLMs express safety instructions (The Safety Operator). Earlier in my PhD I studied learning dynamics and inductive biases, using iterative linearisation and the implicit effects of learning rates to explain how deep networks learn features.
I also like building things that work at scale. I spent 2.5 years as a software engineer at Google on public transport routing for Google Maps, and before that worked on combinatorial optimisation (Best Student Paper at CP 2017), represented Australia at the IOI, and competed at the ICPC World Finals.
I submitted my PhD at Cambridge, supervised by Hong Ge and advised by Carl Rasmussen in the Machine Learning Group, supported by the Harding Distinguished Postgraduate Scholars Programme. My honours thesis was supervised by Michael Thielscher and Alan Blair, and before that I completed research placements with Andreas Schutt and Haris Aziz.
I’m looking for research roles starting in early 2027. Feel free to get in touch.
Publications
Teaching
- 2024 Co-supervised Varun Jain (MPhil)
- 2024 Lent Term: Supervisor for 3F8: Inference
- 2023 Lent Term: Supervisor for 3F8: Inference
- 2018 semester 2: Teaching Assistant for COMP9444: Neural Networks
- 2018 semester 2: Teaching Assistant for COMP2521: Data Structures and Algorithms
- 2018 semester 1/2017 semester 1: Teaching Assistant for COMP3411: Artificial Intelligence
- 2017 semester 2: Teaching Assistant for COMP1521: Computer Systems Fundamentals
Service
- Reviewer for ICML 2026 — awarded Gold Reviewer
Talks
- 2024: Invited talk on iterative linearisation, MediaTek
- 2022: Invited talk on iterative linearisation, Huawei
Personal
I’ve been juggling since 2006 and hold two world records (15- and 17-ball bounce passing). You can see some related projects here.