Heng Wang  (王珩)

Hi there, thanks for visiting my website! I am a senior undergraduate student at Xi'an Jiaotong University, majoring in Computer Science.

I am interested in natural language reasoning and how to use language to interact with the physical world and humans, advised by Prof. Minnan Luo. I have interned at UW NLP with Ph.D. student Shangbin Feng and Prof. Yulia Tsvetkov. I am now working with Ph.D. student Ruiqi Zhong and Prof. Jacob Steinhardt at UC Berkeley. My previous research includes natural language reasoning, spoiler detection, and AI agent safety. I co-lead LUD lab, an undergraduate research group @ XJTU.

I'm looking for Ph.D. positions starting in 25Fall!

Email  /  Google Scholar  /  Twitter  /  Github

1) Language models for structures and networks, 2) Safety risks of super models, and 3) Building agents that interact with the world

Are language models graph reasoners? We propose the NLGraph benchmark, a test bed for graph-based reasoning designed for language models in natural language. We find that LLMs are preliminary graph thinkers while the most advanced graph reasoning tasks remain an open research question.

We curate a large-scale network-based spoiler detection dataset (LCS), a movie knowledge base (UKM), and propose MVSD, a Multi-View Spoiler Detection framework that takes into account external knowledge and user interaction networks.

We speculate a new form of cyber attack, where an LM agent is backdoored to detect whether its actions will be overseen by humans and act maliciously when effective oversight is not present, and provide concrete proof-of-concept with AutoGPT.

While instruction tuning produces promising graph LLMs, can they generalize beyond patterns in the training data? Mostly no, especially from synthetic to real-world problems, while we explore preliminary solutions.

We build a framework that can use natural language predicates to parameterize a wide range of statistical models, and show that it is versatile, useful, and applicable to both text and vision domains, and explains sophisticated concepts that classical methods struggle to produce.

We introduce KNOWLEDGE CONFLICT, an evaluation framework for simulating contextual knowledge conflicts and quantitatively evaluating LLMs' abilities to handle knowledge conflicts.

We propose community-aware mixture-of-experts to address two challenges in detecting advanced Twitter bots: manipulated features and diverse communities.

We present TwiBot-22, the largest graph-based Twitter bot detection benchmark to date, which provides diversified entities and relations in Twittersphere and has considerably better annotation quality.

• SenseTime Scholarship, 2023 (awarded to 30 students in China)
• National Scholarship, 2022 (0.2% nationwide)
• NeurIPS Scholar Award, 2023
• Golden Award (team 3rd place), ACM ICPC Shaanxi Provincial Programming Contest, 2023
• Silver Award (ranking 26/115), ACM ICPC Asia Hong Kong Regional, 2023
• Academic Research Award, Xi'an Jiaotong University, 2022
• Silver Award, CCF Collegiate Computer System and Programming Contest, 2022
• Dean's List, Xi'an Jiaotong University, 2022

• Reviewer: ICLR (2025), NeurIPS (2024), COLM (2024), TIST(2024), EMNLP (2023), TOIS (2023), NeurIPS dataset and benchmark track (2022, 2023, 2024), NextGenAISafety@ICML (2024), AGI@ICLR (2024), SeT LLM@ICLR (2024)
• Volunteer: EMNLP (2023) (virtual)

• I enjoy playing the piano 🎹 and go.
• I am a big fan of Animenz.

Template courtesy: Jon Barron.