Chapter 28 — Violation of Expectation: Is the Model Surprised by “Impossible Events”?

The big picture

A tiny story

A ball disappears behind a screen and appears across the stage. A child gasps; a motion detector also fires. Both responded, but only one response might involve a physical expectation.

A Violation-of-Expectation (VoE) test makes an expected continuation, intervenes, and measures prediction–observation mismatch. A larger teleport spike first means the tested event disagreed more with the model’s prediction. It does not yet mean the model knows teleportation is impossible.

The real rule

Use paired triplets with the same prehistory, actions, intervention time, and evaluation window:

1. unperturbed continuation;
2. appearance intervention;
3. state discontinuity such as teleportation.

LeWM v3 reports this design in three environments:

• TwoRoom: agent color change versus agent teleport;
• PushT: block color change versus teleporting both agent and block;
• OGBench-Cube: cube color change versus cube teleport.

PushT is visibly not a one-factor “color versus physics” test: the teleport branch changes two objects and their relation, while the color branch changes one object’s appearance.

The paper reports significantly higher surprise for teleportation across the three environments under a paired test with threshold below 0.01, while the main figure’s color effects are weaker and non-significant. The careful wording is: LeWM’s mismatch signal was more sensitive to these teleport interventions than to these color interventions under the authors’ protocol.

The frozen public repository does not contain enough VoE code to recover the authors’ exact reduction. A course implementation may define a tutorial scalar—the mean squared coordinate difference between predicted next latent and the frozen encoder’s latent for the actual next observation—but must label it tutorial diagnostic.

The trick that can fool us

A last-frame predictor copies the current latent forward and ignores actions. Ordinary slow motion gives modest error. Teleportation gives a large spike. A tiny color change gives a smaller spike. It passes the weakest ordering without learning contact dynamics.

Add controls that attack this explanation:

• same actions after every branch;
• overlapping image-change magnitude;
• in-support fast motion and support-matched reset;
• one-object and two-object interventions with matched displacement;
• last-frame, stagnant-latent, action-shuffled, appearance-only, and persistence baselines.

Inspect the full time trace. A response before intervention suggests leakage or alignment error. A one-step spike followed by recovery differs from a lasting plateau, but neither diagnoses its own cause. New observations can re-anchor prediction; persistent error can mean unsupported reset or self-fed drift.

Pre-register intervention strengths, support bins, mismatch reduction, fixed-time or peak statistic, exclusion rules, and episode-level unit. Do not fit normalization after viewing teleport test spikes.

Experiment receipt

Archive the last clean history, action block, intervention frame, before/after observations, affected objects, simulator reset states, policy, random seed, checkpoint, attachment point, and aggregation rule. Keep dashed predicted latents distinct from solid encoded observations. A decoder image is a separate readout unless the protocol explicitly defines its error.

The later ACPC v1 gives the same action sequence to clean and visually perturbed histories and measures rollout divergence. It introduces Invariance Radius (IR) and Separation Rate (SR), with samplewise bounds connecting divergence to multi-step prediction error and—under a shared candidate pool—to planner-cost change. Its authors report trends on four tasks with three seeds per condition and public code 90d4276. The screen uses observed source-task success labels and tests one blur/resize severity; it does not show that selecting plans with ACPC improves control.

VoE supports perturbation sensitivity under a named protocol. It does not by itself prove object permanence, causal physics, human-like cognition, planner use, or generalization to unseen violations.

Quick check

1. Which parts of a VoE triplet must be paired?
2. Why is PushT color versus teleport not a single-factor physics test?
3. Which simple baseline can also be “surprised” by teleportation?