acceptodds
Under review as a conference paper at ICLR 2027

Low-Rank Dynamics-Effective Latent Carriers for Counterfactual Rollout in Learned World Models

Abstract

We ask whether a compact, 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 world model and construct candidate intervention carriers from training-only counterfactual-minus-factual hidden differences. 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 velocity edits, rank 4 is the smallest tested rank satisfying the registered development criteria. A one-shot rank-4 patch launches a 12-transition autonomous rollout, and the frozen procedure satisfies a preregistered 2-of-3 fresh-checkpoint replication rule and remains reusable at nearby anchors. The same Single-derived carrier and Single-only map also support bounded same-object Joint requests. Across matched training regimes, broader counterfactual intervention support is associated with better Joint rollout fidelity and more additive hidden responses. Carrier-relative diagnostics show that the rank-4 subspace is not dynamically closed: its effect couples strongly to the wider recurrent state. These results support a compact intervention-entry interface rather than an intrinsic four-dimensional state, and suggest addressable latent interfaces as a potential design target for controllable world models.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
World Models, Latent Interventions, Low-Rank Intervention, Counterfactual Dynamics, Mechanistic Interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ocoAOpenReview ↗
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
  1. 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?
  2. 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.
Reviewer_qKpAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies whether the intervention on a recurrent world model hidden-state can lead to an intended counterfactual trajectory and then let the model continue it. The authors study this question using a controlled two-object collision environment with a 192-dimensional recurrent world model.

Critical strengths

S1: The paper propose a technic to design controllable world models, using hidden state intervention for world model control.

Critical weaknesses

W1: The paper performs its study on a single task, a controlled two-object collision environment. The experiments are also limited to a simple 192-dimensional recurrent world model.

W2: I personally found the paper difficult to follow, using terms before or without necessary context/definition. I think the main paper would benefit a lot from adding a simple figure to visually explain the method, and also for the two-object collision environment. I also think that terms like "Single", "Joint" or "Carrier" should be defined or contextualized before being used in the abstract.

W3: Comparison with other methods/technics for world model control: The paper explore a technic for world model control but did not compare or relate to existing technics like latent actions-based world models [1].

[1] Garrido, Quentin, et al. "Learning latent action world models in the wild." Forty-third International Conference on Machine Learning. 2026.

Questions for authors

Q1: Did you further study the effectiveness of your method on other simple environments like Maze, Reach, Wall, Push-T [2] or more complex and commonly used environments: DeepMind Control tasks, Meta-World/ManiSkill2, Atari games, ...

[2] Zhou, Gaoyue, et al. "Dino-wm: World models on pre-trained visual features enable zero-shot planning."

Reviewer_xs1GOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether a low-dimensional intervention in a world model's hidden state can produce a desired counterfactual trajectory that remains consistent over future predictions. The authors use SVD to identify a low-rank subspace from factual and counterfactual hidden-state differences in a GRU-based world model, and train a linear mapping to predict the required latent intervention from the physical state and requested velocity change. Experiments in a simulated two-object collision environment show that a rank-4 intervention enables 12-step autonomous counterfactual rollouts, generalizes to combined velocity edits, and satisfies the evaluation criteria on two of three independently trained models.

Critical strengths

S1. The paper studies an interesting and relatively novel problem. I think exploring low-rank interventions in world models is a meaningful research direction.

S2. The proposed method seems to work in the tested setting. The results show that a low-rank intervention can successfully guide counterfactual rollouts, which is an interesting finding.

S3. The observation that the low-rank subspace interacts with the remaining hidden states is also interesting. It provides some insight into how these interventions affect the model's dynamics.

Critical weaknesses

W1. The experimental setting is too simple. The paper only studies a small world model in a 2D environment with two objects, using numerical position and velocity inputs. I am not sure whether these findings can generalize to more realistic and complex world models. This is my biggest concern about the paper.

W2. The main finding is that a rank-4 subspace can control counterfactual rollouts in a 192-dimensional hidden state. However, the original physical system only has 8 state variables, with 4 editable velocity components. Given this simple setting, the rank-4 result is not very surprising. I am not sure how much new insight this provides about the latent dynamics of world models.

W3. The paper uses SVD and ridge regression to construct low-rank interventions, but the benefit of the low-rank design is unclear. A simple baseline would be to directly predict the full 192-dimensional hidden-state difference using the same training data and ridge regression, without SVD. Comparing rollout accuracy, success rates, and Joint Edit performance would help show whether the low-rank design actually provides an advantage.

Questions for authors

Q1. The experiments are limited to a very simple 2D environment with two objects and a small world model. Could you provide evidence or discuss that the method also works in a more realistic or complex setting, such as a larger video world model?

Q2. The physical system has only four editable velocity components. How can we tell whether the rank-4 result reveals meaningful latent structure rather than simply reflecting the low-dimensional physical system?

Q3. Could you compare your method with a simple baseline that directly predicts the full 192-dimensional hidden-state difference using ridge regression, without SVD? Would the low-rank method still provide an advantage in rollout accuracy or success rate?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.