AI & ML Research 2025
AI False-Belief Tracking
Belief-sensitive embodied assistance
The problem
An embodied assistant that only tracks the true state of the world will happily hand someone keys from the wrong room, or stay silent when a person's plan is built on a mistaken belief - it has no model of what the human actually thinks is true.
The approach
Built an agent with an explicit belief-inference module (what does this person think is true, and why) layered under an intervention-decision network that chooses, per situation, whether to physically act on the human's behalf or correct their belief through communication.
This project studies embodied agents that model Theory of Mind - specifically, agents that can detect when a human holds a false belief about the state of the world, and decide what to do about it.
the classic false-belief problem, embodied
The canonical example: a person believes their keys are in the kitchen, but they were actually moved to the bedroom. A purely world-state-aware agent either does nothing (not its problem) or silently “fixes” the situation by fetching the keys itself - neither of which is guaranteed to be the right call. Sometimes the right move is to say “your keys are actually in the bedroom,” not to act unasked.
architecture
The system is built in two layers on top of the VirtualHome simulator:
- Belief-inference network - maintains a model of what the human believes about object locations and world state, distinct from ground truth, updated from what the human has and hasn’t observed.
- Intervention-decision network - given a detected belief/reality gap, decides whether the right response is to physically act (move the object, complete the task) or communicate (tell the human the truth) - or do nothing at all.
belief = belief_model.infer(human_observation_history)
gap = detect_divergence(belief, world_state)
if gap:
action = intervention_policy.decide(gap, task_context)
# action in {"act", "communicate", "abstain"}why VirtualHome
VirtualHome’s household simulator generates false-belief situations for free - objects get moved, doors get closed, states change while an agent (or a human proxy) isn’t looking. That makes it a natural testbed for training and evaluating belief-inference and intervention policies without hand-authoring every false-belief scenario.
status
This is an active research project. The core belief-inference and intervention-decision components exist; ongoing work is on evaluating intervention policy quality across a broader set of household scenarios.
What I took away
- Tracking the true world state is the easy half; inferring what a human believes - and when that belief diverges from reality - is where the real modeling problem lives.
- Helping isn't always the right action: sometimes the correct intervention is telling someone they're wrong, not silently fixing it for them.
- VirtualHome's household simulator is a good testbed precisely because false beliefs (an object moved, a door closed) arise naturally from ordinary state changes.
0 comments
Loading…