zhengxuan wu

i go by zen, a CS Ph.D. student at Stanford University, advised by chris^2: . i'm a research fellow at Transluce. and i spent a few years on learning a few things about space combustion and cloud computing.

i'm now interested in interpreting language models and, thus, controlling them.

selected works (*=equal contribution)

Language model circuits are sparse in the neuron basis

Blogpost, 2025

Aryaman Arora* · Zhengxuan Wu* · Jacob Steinhardt · Sarah Schwettmann

Improved representation steering for language models

NeurIPS, 2025 (spotlight)

Zhengxuan Wu* · Qinan Yu* · Aryaman Arora · Christopher D. Manning · Christopher Potts

code · blog

AxBench: Steering LLMs? Even simple baselines outperform sparse autoencoders

ICML, 2025 (spotlight)

Zhengxuan Wu* · Aryaman Arora* · Atticus Geiger · Zheng Wang · Jing Huang · Dan Jurafsky · Christopher D. Manning · Christopher Potts

code · viz

ReFT: Representation finetuning for language models

NeurIPS, 2024 (spotlight)

Zhengxuan Wu* · Aryaman Arora* · Zheng Wang · Atticus Geiger · Dan Jurafsky · Christopher D. Manning · Christopher Potts

code · blog

pyvene: A library for understanding and improving PyTorch models via interventions

NAACL, 2024

Zhengxuan Wu · Atticus Geiger · Aryaman Arora · Jing Huang · Zheng Wang · Noah Goodman · Christopher D. Manning · Christopher Potts

code

Interpretability at scale: Identifying causal mechanisms in Alpaca

NeurIPS, 2023

Zhengxuan Wu* · Atticus Geiger* · Christopher Potts · Noah D. Goodman

code · blog

zhengxuan wu volume_up Your browser does not support the audio element.

selected works (*=equal contribution)

zhengxuan wu