HUMAIN Lab

HUMAIN: HUman-centered MAchine INtelligence

We are a group of researchers passionate about AI and sometimes worried about it. We think deeply about how to build controllable machine intelligence that works in the best interests of people.

What do we believe in?

Machine intelligence should be useful, controllable, and understandable, especially when it interacts with humans at scale. Capability alone is not enough; controllability and interpretability matter just as much.

What are we thinking about these days?

Foundations

We study how modern machine learning systems (especially large language models) generalize, and when they may fail.

A key question we are currently exploring is disentanglement. Learning at today’s scale is deeply interleaved: we cannot easily separate what a model has learned from what it has not, nor can we reliably determine how different capabilities extrapolate. We aim to develop methods to disentangle the learning process itself, analyze what trained models have actually learned (e.g., separating knowledge from skills), and design training procedures that are disentangled from the ground up.

Personal AI assistants

The north star of our lab is to build a caring and capable personal AI assistant (e.g., Baymax). To achieve this, we focus on training methods that improve core capabilities such as:
•Learning to learn: We cannot teach a model everything at training time. An ideal assistant should be able to learn effectively at test time and extrapolate safely from limited learning signals.
•Lifelong learning: How should knowledge be encoded and updated? How should memory and long-context management work, especially when the goal is to represent a user’s goals, preferences, and evolving life context, while respecting privacy?
•Understanding of people: How can we anticipate users’ needs and enable meaningful interactions between the user and the assistant?
Many of the concrete problems we work on are related to mid/post-training, including:
•Efficient RL post-training algorithms with or without verifiable rewards, e.g., for reasoning, alignment, collaborating with humans, and multi-agent collaboration.
•Reverse engineering the skills and knowledge a model has learned during training.
•Architectures for modeling and simulating people and for enabling personalization.
•Monitoring and intervention tools to prevent reward hacking.

How do we work?

We value curiosity and believe research should be fun and fulfilling. Work-life balance matters. We are question-driven and learn whatever we need (technical or not) to answer the questions we care about.

Prospective students

If you are interested in joining us, please fill in the Google form.