Compact but Moving: Intervention-Relevant Geometry in Recurrent World Models

We ask whether a small, directly addressable hidden-state intervention can place a learned world model on an intended counterfactual future and then let the model's own dynamics carry that future forward. In a controlled two-object collision environment, we study a 192-dimensional recurrent model trained on factual and locally edited counterfactual trajectories. Candidate carriers are learned from training-only counterfactual-minus-factual hidden differences, and an affine map predicts carrier coordinates from the factual state and requested edit without access to the native counterfactual hidden state at test time. For bounded single-component velocity edits, rank 4 is the smallest tested rank on the preregistered grid that satisfies the development criteria. A one-shot rank-4 patch launches a 12-transition autonomous rollout without future observations, teacher forcing, repeated hidden-state correction, or physical-state clamping. The frozen procedure satisfies the preregistered 2-of-3 fresh-checkpoint replication rule and remains reusable at nearby anchors. The same Single-derived carrier and Single-only affine map also support bounded same-object two-component requests. Across the matched training regimes, broader counterfactual support was associated mainly with better Joint rollout accuracy and more additive Joint hidden responses. Composition-related structure is enriched in the rank-4 subspace but is not confined to it, and local recurrent diagnostics show strong one-step coupling from the carrier to the rest of the hidden state. A position-edit stress test fails the required specificity controls. Together, these results support a compact dynamics-effective intervention-entry interface, not a closed four-dimensional state or an intrinsic state dimension.

Interpretability methods such as probes, activation patches and learned editors are designed to reveal or modify a model's current computation. World models pose a harder requirement: because their predictions become inputs to later predictions, a useful internal correction must survive after editing stops. We therefore propose RolloutFaith, a framework that measures semantic improvement both in the prediction produced at intervention time and over later autonomous predictions under fixed events, actions, noise, and information budgets. We evaluate ten fitted editors on three world models across Crafter, Cartpole, and CoinRun. We also use Reference Activation Patching, which replaces a model activation with the paired activation computed from the real observation, to measure the correction available at the chosen interface. This reference intervention improves later predictions in all nine model and task combinations and outperforms the best fitted editor in eight, yet its sustained gain decreases with horizon in five of nine combinations. Current fitted editors recover only limited and inconsistent long term effects. By restoring individual state components to their untouched values, we find that persistent effects travel through the newest generated frame in DIAMOND, recurrent memory in DreamerV3, and both in STORM. These findings suggest that training should reward future consequences. To test this hypothesis, we propose Delayed LoReFT, which optimizes the same low rank intervention through four frozen future transitions and improves sustained intervention effects to some extent.

Planning-time intervention fidelity is a distinct, measurable property of a learned world model: whether the model's own open-loop transitions move task variables the way matched environment interventions do. In the settings we test, it is neither revealed by reward fit nor ensured by task-anchored training. Across released TD-MPC2 checkpoint sizes, episode return falls as an operator-error diagnostic on task observables grows, while reward-prediction error stays small and nearly flat, and a self-supervised world model trained without task signal preserves the same operator substantially better than a task-anchored model on the shared task. A capture-gated matched-intervention audit then localizes what fails. On Cheetah, three LeWorldModel checkpoints capture the current task query and support decodable real intervention effects; however, their imagined five-step effects are worse than predicting no effect and worse than an environment-endpoint oracle. The failure is task-direction rotation with excess gain, not feature collapse. This severe pattern is conditional: five PreJEPA seeds retain an oracle-relative deficit without it, Finger Spin experiments extend the deficit beyond locomotion with heterogeneous severity across seeds, and shared-bank effect geometry is both candidate- and support-dependent. We also test practice-side questions. In DreamerV3 the posterior distribution, not its sample, carries the current query; ensemble disagreement ranks error only near training support; and a frozen support-aware score degrades held-out error ranking in both tested transfer directions while native disagreement remains informative in both. We conclude that intervention fidelity must be audited directly, capture-first, on the model's native interface.