Chapter 19 — Long-Horizon Rollouts: How Small Errors Snowball into Major Failures

The big picture

A tiny story

Make one photocopy of a picture and it may look fine. Copy the copy many times and small distortions can matter.

LeWM rollout is autoregressive. The first prediction starts from encoded real observations. Later predictions can consume earlier predicted latents. Training and one-step tests usually see more real contexts than a long planning rollout does. This is the train–rollout context gap.

The photocopy story has a limit: error need not grow smoothly or forever. Dynamics can amplify, rotate, cancel, or damp a discrepancy. Measure the curve; do not assume exponential or monotonic drift.

The real rule

Keep two lanes for the same held-out episode and the same recorded action sequence:

• Recorded-context lane: encode the real observation again at every step.
• Self-fed lane: after the initial history, feed predictions back into the next prediction.

Compare each predicted latent with the frozen encoder’s latent for the observation that actually followed. Plot error by model horizon, not only at the endpoint. Slice by door crossings, contact onset, contact loss, action magnitude, and empirical support.

Original LeWM v3 emits one latent estimate. It does not output a calibrated confidence or a distribution of possible futures. Deterministic output does not mean the environment is certain.

One-step accuracy is still useful. It simply answers a smaller question. Good recorded-context prediction can coexist with poor self-fed rollout, a bad terminal cost, or a broken execution adapter.

The trick that can fool us

CEM is paid to find low predicted cost. If one unsupported action sequence makes the learned rollout unrealistically optimistic, more candidates or refinement rounds can discover that loophole more reliably.

Test this claim instead of naming every failure “model exploitation”:

1. Freeze a bank of action sequences.
2. Compare imagined cost with real executed outcome.
3. Label action-support and horizon.
4. Increase search budget while keeping the model, task, and real-action budget fixed.

Exploitation becomes plausible if predicted best cost improves, selected sequences move farther from support, and real outcomes worsen in a paired way. A final crash alone cannot separate dynamics, cost, search, support, and execution.

If oracle cost fixes candidate ranking but rollout endpoints remain wrong, repair dynamics. If oracle dynamics plus learned cost still prefers the wrong route, repair cost or representation geometry. Shortening the horizon reduces rollout depth; it does not fix a ruler that ranks across-wall states incorrectly.

Experiment receipt

Store the checkpoint, preprocessing, initial real history, aligned recorded actions, action-block semantics, teacher-forced and self-fed latents, per-horizon errors, support labels, and all seeds. Mark the first qualitative divergence, not only the last ugly frame.

The mechanism is grounded in frozen LeWM rollout code at author commit 8edfeb3 and the limitation discussion in LeWorldModel v3. The model-exploitation idea also has neighboring precedent in Model-Based Planning with Energy-Based Models.

The independent TwoRoom reimplementation reports that one-step error orders short-horizon but not long-horizon success across three checkpoints in one environment, with one training seed per configuration. It leaves the cause open; it does not prove universal drift or exploitation.

Quick check

1. What changes between recorded-context and self-fed prediction?
2. Why is one deterministic latent not calibrated certainty?
3. Which paired trend would make model exploitation a credible hypothesis?