- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper examines whether a compact hidden-state edit can put a recurrent world model on counterfactual rollout and let the model's own dynamics carry it forward. The setting of the experiment is a 192-D recurrent world model of two-object collision environment. The authors built a low-rank carrier from SVD of hidden differences between counterfactual and factual state. The authors interpret rank 4 as a compact intervention entry interface into the model's high-dimensional recurrent dynamics, not an intrinsic or closed four-dimensional state. Rank 4 is the smallest tested rank that passes the registered Single-edit assay, and almost all of its one-step effect couples into the rest of the hidden state rather than staying inside the carrier.
Critical strengths
- Compare to prior steering and probing work, this work sets a stricter evidence standard. It requires a hidden edit must be applied once and then produce the intended 12-step rollout under the model's own dynamics.
- The paper cleanly separates subspace capacity, measured by an oracle projection, from addressability, measured by a patch predicted without access to the counterfactual hidden state.
Critical weaknesses
- The main findings of the paper appear to follow from the architecture. Observations enter the hidden state only through each object’s 16-D middle component, so the native hidden differences live in at most 32 of 192 dimensions. A four-dimensional request space (2 objects x 2 velocity axes) then predicts a rank-4 carrier under local linearity. The middle component is also recomputed by the prior at every step, which might explains the near-zero carrier retention (0.016, below the 4/192 random baseline).
- There are no comparisons to existing intervention methods, such as DAS, probe- or decoder-derived directions, or a simple Jacobian patch of the anchor update. It is therefore unclear what the SVD-plus-ridge construction contributes beyond a learned approximation of a local linearization.
- The evidence remains limited in scope and statistical strength. The evaluation uses a single small architecture, a deterministic simulator with two objects and numeric observations. No confidence intervals are provided.
Questions for authors
- Since the middle component is recomputed by the prior at every step, is near-zero carrier retention expected from the architecture alone? How does the observed retention (0.016) compare with the 4/192 random-subspace baseline?
- I would suggest for future work, compare against existing intervention methods, such as DAS, probe- or decoder-derived directions, and a Jacobian patch of the anchor update.