Kaiqu Liang

I am a 3rd year PhD student in Computer Science at Princeton University, advised by Jaime Fernández Fisac. I also work closely with Tom Griffiths.

Previously, I completed my MPhil in Machine Learning at the University of Cambridge, advised by Samuel Albanie and Bill Byrne. I did my undergraduate studies at the University of Toronto, where I was advised by Roger Grosse and Sven Dickinson. During my undergraduate years, I was also a student researcher in Vector Institute, advised by Roger Grosse.

Email  /  Google Scholar  /  Twitter  /  Github  /  Linkedin

News

Research

I'm working on human-AI safety, LLM planning & alignment, and embodied AI. My long-term goal is to build safe, reliable AI systems that effectively assist humans. Previously, I also did research on video retrieval and out-of-distribution generalization. I am happy to collaborate and answer questions about my research.

- I've been considering these questions on human-AI safety & alignment: Are we underestimating the ways LLMs might subtly influence daily human decisions? Do we recognize that certain harms might surface long after the interaction? How should our alignment algorithms account for these risks?

- I've been considering these questions on embodied intelligence: LLMs are powerful for reasoning, but are they sufficient for robot planning? What are their capabilities, when should we use them, and how can we augment them with other tools?

We found that RLHF can induce significant misalignment when humans provide feedback while implicitly predicting future outcomes, creating incentives for LLM deception. To address this, we propose RLHS (Hindsight Simulation): By simulating future outcomes of the interaction before providing feedback, we drastically reduce misalignment.

We propsoed introspective planning as a systematic approach that aligns LLM’s uncertainty with the inherent ambiguity of the task.

We introduced Branch and Play (B&P), an algorithm that effectively resolves multi-agent spatial navigation problems by determining the optimal order of play.

We proposed several simple yet effective baselines for interactive video retrieval via question-answering.

We demonstrated that equilibrium model improves generalization in harder instances due to their path independence, highlighting its importance for model performance and scalability.

We demonstrated and discussed why Deep Equilibrium (DEQ) Models outperform fixed-depth counterparts in generalizing under distribution shifts.

Education

Teaching

Reviewer services

Website source from Jon Barron