shu.b@northeastern.edu
Hi! I am Bangzhao Shu (舒邦照), currently a second year Ph.D. student in Computer Science at the Khoury College of Computer Sciences at Northeastern University, advised by Professor Mai ElSherief.
Broadly, my research lies at the intersection of human-centered NLP and LLM safety. My current research focuses on the following themes:
(1) Emotional and Social Intelligence in LLMs: Human communication is deeply shaped by emotions and social norms. I study how language models perceive, reason about, and respond to emotional and social contexts, and develop evaluation frameworks to assess their behavior.
(2) Mechanistic Interpretability of LLMs: Large language models are often treated as black boxes, making it difficult to understand and control their behavior. I investigate the internal mechanisms that give rise to model behavior, with the goal of improving interpretability and reliability.
(3) Alignment and Safety of LLMs: Ensuring that model behavior is reliable, safe, and aligned with human values is critical. I study methods for evaluating and improving alignment, including identifying failure modes and developing interventions to guide model behavior.
Before starting my Ph.D., I earned dual master’s degrees in Information and Environment & Sustainability from the University of Michigan. I worked under the guidance of professor David Jurgens at UMSI and professor Yuhao Kang at the University of Texas, Austin.
Education
Ph.D. in Computer Science, Northeastern University, Expected 2024–2029
M.S. in Information, University of Michigan, 2022–2024
M.S. in Environment & Sustainability, University of Michigan, 2021–2024
B.E. in Geomatics, Nanjing Normal University, 2016–2020
Publications
You don't need a personality test to know these models are unreliable: Assessing the Reliability of Large Language Models on Psychometric Instruments
Bangzhao Shu*, Lechen Zhang*, Minje Choi, Lavinia Dunagan, Dallas Card, David Jurgens
Proceedings of the 2024 Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL 2024)
Fluent but Unfeeling: The Emotional Blind Spots of Language Models
Bangzhao Shu*, Isha Joshi*, Melissa Karnaze, Anh C Pham, Ishita Kakkar, Sindhu Kothe, Arpine Hovasapian, Mai ElSherief
The 20th International AAAI Conference on Web and Social Media Understanding the World Through the Web (ICWSM 2026)
Causally Modeling the Linguistic and Social Factors that Predict Email Response
Yinuo Xu*, Hong Chen*, Sushrita Rakshit*, Aparna Ananthasubramaniam*, Omkar Yadav*, Mingqian Zheng*, Michael Jiang*, Lechen Zhang*, Bowen Yi*, Kenan Alkiek*, Abraham Israeli*, Bangzhao Shu*, Hua Shen*, Jiaxin Pei*, Haotian Zhang*, Miriam Schirmer*, David Jurgens
Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL 2025)
"He gets to be the fun parent": Understanding and Supporting Burnt-Out Mothers in Online Communities
Nazanin Sabri, Ananya Malik, Bangzhao Shu, Jason Snyder, Laurie Kramer, Mai Elsherief
arXiv
Teaching
Teaching Assistant: CS 6120 Natural Language Processing, Northeastern University, Winter 2025
News
July 2025: Our new paper Fluent but Unfeeling: The Emotional Blind Spots of Language Models was accepted to ICWSM 2026! We benchmark a fine-grained emotion recognition dataset EXPRESS with 251 unique emotion labels. We systematically tested LLMs and found that it remains challenging for them to align with human on emotion nuances.
May 2025: Our new paper Causally Modeling the Linguistic and Social Factors that Predict Email Response was accepted to NAACL 2025!
Sep 2024: Started my new journey at Northeastern University!
May 2024: I graduated with a Master of Science in Information and a Master of Science in Geospatial Data Science from the University of Michigan!
April 2024: Our paper You don’t need a personality test to know these models are unreliable: Assessing the Reliability of Large Language Models on Psychometric Instruments was accepted to NAACL 2024! In this work, we construct a dataset covering 39 psychometric instruments across 115 persona axes. We show that current question-style prompts are insufficient to capture model perceptions reliably -LLMs are highly sensitive to simple perturbations and exhibit low consistency under negation.