Please select your name from the dropdown list after filling out the registration form.

The Bayesian theory-of-mind model uses a POMDP solver to compute what actions A a rational agent should given their current belief B_t and desire D. However, POMDP solvers typically result in plans / policies that are deterministic (i.e. a single optimal action is taken at each belief state).

How does the BToM model turn this into a distribution over actions instead, P(A | B_t, D)?

Is this a reasonable probabilistic model of how agents select actions?

As a baseline, the BToM model is compared against a cue-based "MotionHeuristic" model which only takes into account which objects the agent is moving towards / away from.

Why is the MotionHeuristic model unable to produce human-like inferences about the agent's desires, using the scenario in Figure 1 as an example?

Why is MotionHeuristic better at producing human-like inferences about the agent's beliefs (as Figure 5d shows)?

How might the scenarios be modified to "break" the MotionHeuristic, so that it no longer produces human-like inferences about the agent's beliefs?

To the extent that LLMs can replicate theory-of-mind associated capabilities like attributing beliefs and desires, do you think they are model-based or cue-based (or something in between)? How could we design experiments to tell?

What is the POMDP solver used for planning in Experiment 2, and how is it related to NUS?