Benedict Aaron Tjandra
I’m a Senior Research Engineer at Google DeepMind in London, working on Gemini pre-training and continual learning.
About
My research spans language models, temporal graph learning, and the theoretical foundations of neural architectures. Recent work includes SkillSmith, which combines learned skills with textual knowledge, and research on the trade-off between information propagation speed and fidelity in computational graphs.
Selected publications
SkillSmith: Learning to Compose Parametric Skills and Textual Knowledge
COLM 2026 LLA Workshop
We combine learned skills and textual knowledge to synthesize new skills through instruction-steered prefix generation.
Feedforward Mixing is as Sharp as it is Slow in Reverse
ICML 2026 FoGen and GFM Workshops
We establish backward mixing time as a bound on information fidelity, revealing a speed–fidelity trade-off in symmetric feedforward computational graphs.
T-GRAB: A Synthetic Diagnostic Benchmark for Learning on Temporal Graphs
KDD 2026 · Data & Benchmarks (Oral)
A synthetic benchmark that isolates temporal reasoning skills and exposes limitations in temporal graph neural networks.
Background
I completed my MSc in Advanced Computer Science at the University of Oxford, where I worked with Yarin Gal in OATML on language-model uncertainty and with Michael Bronstein on temporal graph learning. Before that, I studied at the University of Cambridge, where I graduated as valedictorian of my cohort. I’ve also held research appointments at Oxford, Cambridge, Mila, and EMBL-EBI, and previously worked in engineering at XTX Markets and Amazon Web Services.
