Zhengxuan Wu

Zhengxuan Wu is a safety researcher at Meta, working on language-model safety and control. He earned a CS Ph.D. at Stanford, where his research explored mechanistic interpretability and behavioral control through interventions on models’ internal representations; he also co-authored pyvene, an open-source library for studying and applying such interventions.

Career

Cosigned byLoading...

CosignedSign in to view

Elsewhere

Stanford University

CS Ph.D.

Products & media

zen-wu.social

Mechanistic Interpretability and Control for Language Models

His dissertation examines causal interventions on language-model representations for interpretability and behavioral control, including DAS, ReFT, AxBench, and RePS.

Writing

aclanthology.org

pyvene: A Library for Understanding and Improving PyTorch Models via Interventions

This paper introduces pyvene, an open-source library for customizable interventions on PyTorch models, with examples in interpretability research.

Writing

Profile name

Company name

Company name

Profile name

Company name

Company name

Profile name

Company name

Company name

Profile name

Company name

Company name

Profile name

Company name

Company name

Zhengxuan Wu

Meta

Stanford University

Products & media

Mechanistic Interpretability and Control for Language Models

pyvene: A Library for Understanding and Improving PyTorch Models via Interventions

Curated Lists

List Inclusions