acceptodds
Under review as a conference paper at ICLR 2027

Event-Aware Tensor-Network Inference for Path Risk in World Models

Abstract

Estimating whether an event occurs along a future trajectory requires information that accurate terminal-state prediction need not preserve. Matrix product states (MPS) can compress this computation for structured stochastic dynamics, but standard singular-value truncation minimizes local error without regard to its effect on the queried probability. We propose event-aware tensor-network inference: an approximate backward survival function weights forward truncation residuals by their effect on the target event probability, and a cost-aware rule reallocates bond dimension across space and time, with backward propagation, allocation, and recomputation charged to the inference budget. The resulting indicators track the true effect of each compression better than residual magnitude, especially under nonlocal coupling (median Spearman 0.68 vs. 0.39), and ablations confirm that both event sensitivity and the cost term contribute. On structured benchmarks with exact references, the method matches fixed-rank accuracy at equal or lower recorded cost, and in a visual path-event proof of concept it improves Brier score under moderate occlusion. Benefits depend on correlation structure and the evaluated scale is limited; exact propagation remains preferable on small systems.

Primary area:
learning on time series and dynamical systems
Keywords:
Tensor Networks, Path-Risk Inference, Event-Aware Rank Allocation, Structured Stochastic Dynamics, World Models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1jLsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes MPS-based inference for estimating path-event probabilities. It combines forward truncation with approximate backward survival functions to allocate bond dimensions according to event relevance and computational cost.

Critical strengths

None

Critical weaknesses
  1. The method assumes finite-state beliefs and structured transition operators, but does not adequately explain how these are obtained from an existing learned world model.

  2. Section 4.3 does not identify the dataset, evaluation size and splits, specific event predicates, or occlusion protocol, making the reported results difficult to interpret or reproduce.

  3. The main comparisons cover small binary-chain systems and selected operating points rather than matched budgets.

Questions for authors
  1. Which visual world model was used, and how were its outputs converted into the finite-state beliefs and structured stochastic transitions required by the method?

  2. What dataset, evaluation splits and sample counts, event predicates, and occlusion-generation procedures were used in Section 4.3?

  3. The paper appears to be missing the AI use statement. Could the authors clarify this omission and provide the required disclosure?

Reviewer_sUaDOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents a technique for matrix product state (MPS) compression that is event-probability-aware, as an alternative to standard L2-based compression, with a view toward successive compressions within a stochastic propagation of a probability distribution encoded as an MPS. The sense in which it is really an alternative is subtle - it comes down to a different choice of truncation threshold than one based on L2 error, so ultimately it only changes how MPS rank is allocated among bonds.

Critical strengths

I like the idea of the task-aware compression. This is the strongest aspect of the paper, and I would want to see this be the focus, tested more rigorously, provided that a compelling domain of application could be convincingly carved out where MPS or more general tensor network structure is more clearly motivated.

Critical weaknesses

The exposition is very difficult to follow, and I think the world model nomenclature is distracting from the actual structure and contribution of the ideas. In my view the presentation is somewhat inverted. The basic picture of the stochastic dynamics and the objects relevant to MPS compression is clarified after the details of a particular mixture model are given, whence the world model nomenclature derives. The mixture model is apparently only relevant for the video example, but the video example is not nearly well-specified. In particular the origin of the belief state b and the stochastic dynamics T is unclear, and the meaning of the x_j variables is not specified.

Moreover, the presentation of the stochastic dynamics seems to rest on an analytical presentation via even-odd gate decomposition, which does not strike me as tractable or relevant in real-world scenarios that the paper gestures toward. I would be skeptical even of a more flexible presentation of T as an MPO. I cannot imagine a pipeline for furnishing b and T that would respect the requirements of the paper’s underlying algorithm. The toy example is better-specified, but in my view the results are small-scale and not too practically inspiring, and I cannot see how the framework extends to more interesting or real-world examples.

Ultimately the singular vectors are the same as the L2-based / naive compression, so the upgrade is all a matter of rank allocation, and it does not seem to reliably pay off in practice. Moreover the experiments are not anywhere close to a computationally limited regime. Even if there is an improvement, I’m not really persuaded that there will exist examples in which the advantage over naive MPS compression is really make-or-break / significant. The more fundamental question in my view is whether the MPS itself is a reasonable choice of representation, i.e., whether or not the entanglement structure of the variables respects low bond dimension.

Questions for authors

N/A

Reviewer_fN2DOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a new inference method with using tensor networks. They allocate the available computational capacity based on the specific event they wish to predict. Instead of pruning representations based solely on local SVD error, the method uses an approximate backward survival function to estimate how much each pruning affects the probability of the event. The method shows some better correlation with the actual prediction error, e.g. in non-local interactions, and achieves similar accuracy at sometimes lower computational cost on benchmarks. It also demonstrates some improvment in prediction of visual events under moderate occlusion. However, the benefits depend on the problem structure, and the evaluation remains relatively limited in scope.

Critical strengths

a. They authors introduce a query-dependent compression in tensor networks, combining ideas from goal-oriented numerical analysis and MPS. b. The mechanism takes into account not only accuracy but also the cost of backward propagation and recomputation c. The work attemps to address a fundamental problem. An optimal representation of the overall state is not necessarily the optimal one for predicting a specific risk or event. d. Benchmarks, ablations, and application to optical trajectories are provited with evidence that event specific sensitivity outperforms the simple truncation residual.

Critical weaknesses

a. The main limitation is that there is limited experimentation and the results are mixed. It is not yet clear whether the benefits hold in much larger systems or under more complex dynamics. b. The method appears to be particularly beneficial when non-local correlations are present in other cases exact propagation remains superior. c. Not clear if the additional complexity of the system matches the imporvments. The estimation of the backward survival function and the dynamic reallocation of bond dimensions may reduce the practical benefit of this method. This is based on accuracy–cost trade-off which is again not 100% clear.

Questions for authors

a. How do the method’s total runtime and memory usage scale in larger systems, when the cost of backward propagation and recalculations is fully taken into account?

b. Can the method be extended to environments with continuous latent states? Especially as the paper opens up with World Vision models.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.