PhDing @MIT | SR @Google, Prev @UCBerkeley. Interested in developing AI and understanding minds.
Joined August 2012
- Tweets9
- Following86
- Followers97
- Likes8
Pinned Tweet
Our paper on novel game reasoning is out in @Nature!
Games occupy a special place in the science of intelligence. While most prior work has focused on acquiring expertise in games, here we study how people can flexibly reason about games that they have never seen before.
Across a series of behavioral studies with 1,000+ participants and 100+ grid-based board games, we find that people are systematic and adaptively rational in their judgments about how fair and fun a novel game is.
We explain this with the “Intuitive Gamer” cognitive model, based on principles of fast, depth-limited, and goal-directed probabilistic simulation. It captures patterns in human data well and offers insights into how people can reasonably evaluate and act when encountering new problems.
This work is also a step towards thinking about AI agents that can not only solve problems but decide which problems are worth solving in the first place.
Exciting release, congrats to the team! Glad to see ideas in our Turing-RL paper echoed in the blog post and user models developed at scale!
arxiv.org/abs/2606.19336
Very inspiring to team up with Katie, who is deeply passionate about games and science!
Games have long served as a testing ground for understanding problem-solving and strategic reasoning in humans and machines, across many fields – from AI, to cognitive science, economics, and more. Yet, in life people are regularly faced not just with repeated instances of the same problem or “system of rule and reward”, but entirely new problems, new systems, new games.
And when encountering new games, people not only need to decide what actions to take but whether a game is even worth playing at all!
How can people think reasonably about new games for the first time, so flexibly and quickly?
We take steps towards this question in our new work, out now in @Nature!
The paper is available here (Open Access). Check out some of our games and see how fair/fun they are :-)
nature.com/articles/s41586-0…
It has been a real privilege to work with an incredible team on this project: @katie_m_collins, Lio Wong, Mau Barba, @gdrtodd_, @adrian_weller, @samisaguy, Tom Griffiths (@cocosci_lab), and Josh Tenenbaum (@MITCoCoSci).
Our new paper on training user simulators! Excited to share this work and think about the many future directions of Turing-RL.
How should we train LLMs for user simulation? A good user simulator shouldn't merely imitate what someone said. It should be indistinguishable from what users could have said.
Introducing 🎭Turing-RL: training user simulators with a Turing-Test reward. The model gets rewarded when an LLM judge can't tell its response from the real users’.
