People, worlds and bodies

Can questions drawn from a hand-written world tell a good world model from a weak one?

A world model is trusted to say what follows from an action and what persists out of sight, yet common video metrics score how output looks. This study asks whether checks built from a hand-written world's fixed canon separate a competent model from a weak one.

Can a world model learn what its actions do, and then lose it?

A world model supports planning only if it carries the consequences of the agent's actions. Predictors that underuse the action have been reported since 2015.

Does a simplified simulation stay faithful when you intervene on it?

Prediction accuracy and fidelity to an intervention's effect are different properties, and a reduced simulation can be trusted for the second on evidence of the first.

What does a learned policy hold still?

A probe shows what a robot policy's internal state contains, not what the policy protects. This study applies Powers's test for the controlled variable to a frozen policy in simulation, asking which scene directions it resists and whether that resistance predicts what it survives.

How compact can a controller be before planning breaks?

Compressing a state to answer the current question need not keep what is needed to update that answer after an action.

Do simulated survey answers echo published results or estimate them?

Agreement with published survey results is the usual evidence that a simulated respondent works. If the model has read those results, the agreement may be recall. This study asks whether exposure to a survey predicts closeness to its results beyond each item's average answer.

What is the smallest real difference a simulated panel can call correctly?

Simulated panels are used to pick which of two designs goes forward, not to predict exact numbers. For that use, the property that matters is resolution, not average accuracy.

Does a simulated population go stale, and how fast?

A simulated population's validity can be claimed once, when it is built, and never rechecked. Two studies show a simulated sample matching real people at one time.

How far can a design be optimised against a simulated panel before real people stop agreeing?

Work on reward models describes how a proxy score parts from real performance as optimisation pressure grows. A simulated panel used to pick the best of many designs is such a proxy.

Do answers that look realistic also match how real people answer?

Judging whether a simulated answer looks realistic is cheap. Comparing it with the real spread of answers is expensive.

Can a simulated person predict how a real individual changes their mind better than a plain statistical model?

Simulated people are mostly judged against group answers, which says little about whether a model tracks one person. Whether person-level validity can be tested with data that already exists decides whether the people side of this area can be tested at all.

Is a person's preference changing, or only the way they express it?

Answers alone may not separate a preference that changes from a fixed preference whose expression depends on history.