01
Diagnose
What does safe-looking behavior leave hidden?
Probe fragile reasoning and internal vulnerabilities that output-based evaluations can miss.
Enyi (Olivia) Jiang
About Me
I study AI safety and alignment through models’ internal representations, developing ways to detect hidden risks and intervene while preserving useful capabilities.
I am a final-year Computer Science Ph.D. candidate at UIUC, advised by Prof. Sanmi Koyejo and Prof. Nancy Amato, and a visiting researcher at Stanford’s Trustworthy AI Research (STAIR) Lab. My broader interests include robustness under distribution shift and applications in climate science and healthcare.
I am also the founder of Inside AI Safety, Inc., an independent nonprofit research organization advancing AI safety from the inside out through representation-level evaluation, robustness, alignment, and intervention. Learn more on our LinkedIn page.
Research · Enyi (Olivia) Jiang
Understanding AI safety
from the inside out.
Internal representations as evidence for evaluation,
signals for alignment, and targets for intervention.
Current & ongoing research
02
Shape
How can representations guide better alignment?
Use latent-space feedback during learning to complement token-level preference objectives.
03
Intervene
When should a safety mechanism step in?
Apply safety mechanisms selectively, balancing protection with useful model capabilities.
Looking ahead · future agenda
Mechanistic safety
Identify the features and pathways that causally contribute to safe behavior in open-weight models.
Representation-level safety evaluation ↗⇄
Agent monitoring & steering
Detect emerging risks across reasoning and action, then redirect agents while preserving legitimate task progress.
CLEAR: selective intervention ↗Selected Research
When should we intervene in a chain-of-thought trajectory to most effectively improve safety?
REACT: Representational Early Analysis and Control of Chain-of-Thought Trajectories
Hanwen Zhang, Enyi Jiang, Ruike Zhu, Qinan Yu, Sanmi Koyejo
arXiv preprint · 2026
REACT applies a single early activation intervention, raising fully safe chain-of-thought and answer outputs from 29.06% to 82.05% on DeepSeek-R1-Distill-Llama-8B while preserving benign behavior.
If latent risk can be detected, can we intervene only when necessary?
CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment
Chengxiao Wang*, Enyi Jiang*, Xiaojing Liao, Sanmi Koyejo (* equal contribution)
Trustworthy AI for Good Workshop @ NeurIPS 2026 · 2026
CLEAR uses a hidden-state gate to adjust a safety adapter’s strength, reducing attack success while preserving useful capabilities.
Does safe-looking behavior necessarily imply a safe internal model state?
When Behavioral Safety Evaluation Fails: A Representation-Level Perspective
Enyi Jiang*, Anders Gjølbye*, Yibo Jacky Zhang, Sanmi Koyejo (* equal contribution)
arXiv preprint · 2026
Models can pass static refusal tests yet remain vulnerable to small internal perturbations, revealing a gap in behavioral safety evaluation.
Education
-
University of Illinois at Urbana-ChampaignComputer Science
Ph.D. Student2023 - present -
University of Illinois at Urbana-ChampaignComputer Science
MS Student (thesis-track)2020 - 2023 -
University of Illinois at Urbana-ChampaignElectrical and Computer Engineering
Undergraduate Student2016 - 2020 -
Zhejiang UniversityZJU-UIUC Institute
Undergraduate Student2016 - 2020
Experience
-
-
Stanford UniversityDepartment of Computer Science
Visiting Student Researcher2026 - present -
MetaResearch InternSummer 2025/2026
The past doesn’t define us. ❤️
May we, by God's grace, create beauty, discover essence, and perceive connections.