Aaquib Syed
Bio
Updated 03/22/26Aaquib Syed is a CS and Mathematics undergraduate student at the University of Maryland, College Park, where he is a Banneker-Key Scholar. He was a fellow in the MATS 5.0 program under Neel Nanda's supervision, conducting mechanistic interpretability research on how refusal is implemented in large language models. His most notable work, "Refusal in Language Models Is Mediated by a Single Direction" (NeurIPS 2024), co-authored with Andy Arditi and others, showed that refusal behavior across 13 open-source chat models is controlled by a single direction in the residual stream. He also co-authored "Attribution Patching Outperforms Automated Circuit Discovery" (NeurIPS 2024 ATTRIB workshop) and work on mechanistic unlearning and model pruning. He is currently a Student Researcher at Google DeepMind on the Frontier Safety team, where he focuses on evaluating and forecasting dangerous AI capabilities.
Community Signal
Updated 03/22/26No endorsements yet.