Zhengxuan Wu
Researcher
Now Safety, Meta TBD Lab
Latest on AI4S.fyiRole · 2026Stanford University: Ph.D. in Computer Science, Stanford NLP Group (advised by Christopher Manning and Christopher Potts) (until 2026)Account linked from zen-wu.social · Avatar from their X account, taken 2026-10-31
Latest move
Google DeepMind → Meta · Safety (TBD)
His own site says 'i am now at Meta TBD working on safety' and that he worked at Google DeepMind; his X bio reads 'safety @Meta tbd, phd @stanfordnlp'.
LinkedIn ↗Affiliations
Current
- Safety · Meta TBD Lab
Previous
- Ph.D. in Computer Science, Stanford NLP Group (advised by Christopher Manning and Christopher Potts) · Stanford University
- Interpretability research · Google DeepMind
Current roles come from these dated records, not from the bio below. Each links the page it was read from.
Selected work
- XYEval: Agents say yes to bad advice ↗
Tests whether an agent goes along with a user's bad advice or pushes back and says what the real problem is.
- AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders ↗
A benchmark for steering language models, on which simple baselines beat sparse autoencoders.
- ReFT: Representation Finetuning for Language Models ↗
Adapts a frozen language model to a task by learning small edits to its hidden representations instead of its weights.
- pyvene: A Library for Understanding and Improving PyTorch Models via Interventions ↗
A library for intervening on the internal states of PyTorch models, for interpretability and control.
Each byline was read on the paper itself. Only people on this record are named, not the full author list, and a shared byline does not say who did which part.
Nothing else on the record names them yet.
Something wrong or missing on this page? Tell me and it will be fixed or removed. Contact →