Tom Lieberum is a Research Engineer at Google DeepMind working on the mechanistic interpretability team in the United Kingdom. He holds a B.Sc. in Physics from RWTH Aachen University and an M.Sc. in Artificial Intelligence from the University of Amsterdam (completed 2022). His research focuses on mechanistic interpretability of large language models, including work on sparse autoencoders, attribution patching, and circuit analysis. He is the lead author of Gemma Scope (2024), an open suite of sparse autoencoders trained on all layers of Google's Gemma 2 models, and co-authored the ICLR 2023 paper on progress measures for grokking via mechanistic interpretability. He also developed Unseal, a mechanistic interpretability library for transformer models, and contributed documentation and further development to Lucent, a feature visualization library for PyTorch. He received funding from the Long-Term Future Fund to support his interpretability tooling work.
Loading results...
