Unfortunately they are student-made slides so my default policy is not to share them!

Hopefully this syllabus is useful to at least some of you!

And if you happen to be in Singapore, I'll be teaching a for-credit version of this class again next semester at NUS, under CS6208! Reach out if you're interested in auditing.

cosilab.notion.site/cs6101-raci-...

The class ends with a discussion of what AI alignment means in a pluralistic world, and -- now that students understand it well enough -- a critique of the descriptive and normative validity of standard theories of rationality.

bsky.app/profile/xuan...

This toolkit can be extended to cooperation at a larger scale: Rational inference allows agents to rapidly adapt to the people, norms & institutions that make up society, while rational deliberation allows agents to move toward social structures that are jointly beneficial.

Rational modeling + decision-making provide a unifying frame for the topics we cover: By modeling other agents as (boundedly) rational, we can infer their goals & beliefs, plan to assist them, work in a team, communicate intent, and teach new concepts.

As the title says, this class focuses on broadly *rational* approaches to cooperative intelligence, grounded in standard theories of epistemic & instrumental rationality (Bayes + utility theory) -- but with some caveats!

Thank you, means a lot!! We were very happy to be able to replicate all the past BToM results :)

Finally, this paper represents the hard work of an incredible team of co-authors:

@lanceying.bsky.social (paper lead), @heyodogo.bsky.social, Katie M Colins, Megan Wei, Ced Zhang, @tbrookewilson.bsky.social, and my co-senior authors Lio Wong + @joshtenenbaum.bsky.social.

Interested in learning more? Catch my coauthor Ced Zhang when he presents LIRAS at #EMNLP2025 in Suzhou this week!

Or find the paper here:
aclanthology.org/2025.finding...

We show that LIRAS correlates strongly with human judgments about an agent's likely goals, beliefs, costs and rewards.

In contrast, large reasoning models like OpenAI o3 show a much weaker correlation, in line with other work indicating that LRMs struggle at ToM reasoning.

As a proof of concept, we tested LIRAS on Theory of Mind reasoning tasks from 7 cognitive science experiments on probabilistic social inference, including key experiments that established the Bayesian theory-of-mind paradigm (Baker et al, 2012/2017; @julianje.bsky.social et al, 2016).

To instantiate LIRAS, we build on recent work by @tbrookewilson.bsky.social suggesting LLMs can serve as ad-hoc model constructors (arxiv.org/abs/2507.12547):
- From language, LLMs synthesize code representing agent + env. models that support coherent inference
- VLMs parse images to env. states

We propose that humans draw such inferences by constructing ad-hoc Bayesian models of agents & their environments, then inferring the agent's goals / beliefs / costs / rewards via inverse planning -- a framework we call Language-Informed Rational Agent Synthesis (LIRAS).

If you stare at these scenarios enough, you'll probably infer that:
1. The robot is helping the human reach gem A
2. The astronaut prefers the orange resource & finds blue terrain dangerous
3. The student thought their favorite foodcart might be behind the building, but was wrong

Humans can rapidly make sense of novel social situations like below:

1. A robot helps a human reach one of 4 gems
2. An astronaut collects resources on the way to their spaceship
3. A hungry student looks for their favorite foodcart behind a building

How do we achieve this?

shockingly, it can both be true that 1) we should improve labor conditions 2) the workers prefer what they have now to what they had before.