acceptodds
Under review as a conference paper at ICLR 2027

Closing the Loop on Vision-Based Control: From GANs to Stochastic World Models

Abstract

Verifying a vision-based neural feedback system requires a model of the observations its controller acts upon. Such a model must capture the variation the sensor produces, while remaining tractable for closed-loop analysis. Generative adversarial networks (GANs) have served as perception surrogates, but they are large, reproduce complex scenes poorly, and are hard to verify. We explore stochastic world models as a richer class of perception surrogates. We train a world model with physically grounded latents, built from operations that standard verifiers bound. It reproduces held-out frames more faithfully than GAN surrogates with up to 130 times as many parameters. To verify these surrogates, we develop a procedure that combines falsification, adaptive refinement, symbolic, and backward analyses. On an emergency braking benchmark with a GAN surrogate, our procedure resolves the entire state space, 38% of which the state-of-the-art verifier left unresolved. On the RGB version of the benchmark, where no verification results have previously been reported, our procedure resolves over 80% of the state space with a world model surrogate.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Formal Verification, Computer Vision, Control Systems

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 4 reviews, above 27% of the others.
Reviewer_d8jGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses verification of vision-based neural feedback systems — closed-loop dynamical systems whose controller observes the state through a camera rather than directly. The key difficulty is that a camera has no closed-form description, so one must replace it with a perception surrogate that maps states to observations. Prior work uses conditional GANs (cGANs) whose latents have no physical meaning, so a verification guarantee over the latent space does not say which environmental conditions it covers; deterministic world models avoid this but have no variation to bound at all.

The paper makes three contributions:

  1. A formalization (Sec. 3) of a vision-based neural feedback system as a state-based system whose sensor is a generative surrogate g : \mathbb{R}^n \times \mathbb{Z} \to \mathbb{R}^d, with the state-based case recovered when g is the identity. This lets state-based verification techniques apply directly.

  2. A stochastic world model (Sec. 4) whose latents are physically grounded — each latent ranges over an interval of a physical quantity, so the possible observations at a state form a box g(s, \mathbb{Z}) of environmental conditions. The architecture is a DCGAN-style deconvolutional decoder with FiLM modulation, built only from layers standard verifiers can bound (linear, conv, batch norm, ReLU, tanh), with the FiLM product handled by a McCormick relaxation.

  3. A verification procedure (Sec. 5, Algorithm 1) combining falsification, forward analysis with abstraction optimization and adaptive refinement, symbolic analysis, and backward analysis, applied cell-by-cell over a partition of the initial set.

Empirically: the world models reproduce held-out frames better than the GAN surrogates on both benchmarks (Table 1) with up to 130× fewer parameters; on the grayscale AEBS benchmark the procedure resolves all 10,000 cells, including the 38% the prior state of the art left unresolved; on the previously unsolved RGB benchmark it resolves over 80% of the state space with a world-model surrogate (Table 2).

Critical strengths

S1. Figure 1 makes the gap visible: the cGAN latent means nothing physically, the deterministic world model has no variation, this one gives a box of grounded latents. "A guarantee over a set of latents does not say which conditions it covers" is the central, correct point.

S2. Folding the sensor into \mathcal{D} = \langle m, n, d, \mathbf{I}, \mathbf{F}, \mathbf{E}, \Xi, h, \mathbf{u}, B, \delta, T, \mathbf{G}, \mathbf{A} \rangle and swapping h for a surrogate g is a small move that pays off: the state-based verification literature (Kochdumper et al., 2023; Sidrane et al., 2022; Akinwande et al., 2026a,b) transfers without re-derivation. Eqs. 2–3 quantify over \Xi explicitly.

S3. Batch norm over layer/group norm, zero-initialised FiLM, McCormick relaxation: each follows from the verification target. 28,591 parameters, 30 s on one H100.

S4. Resolving 100% of the grayscale benchmark where Cai et al. (2025) leave 38% unresolved is a checkable gain. Table 5: 2,208 / 551 / 3 / 3,774 cells decided by forward, abstraction, symbolic and backward analysis; backward alone decides the 2,669 cells the baseline verifies.

S5. Table 1's protocol is right, the SSIM gap (0.713/0.823 against 0.449/0.489) is wide enough, and Figure 4 shows the GANs keep mean intensity and lose structure.

S6. Three real limitations: the guarantee is for the surrogate, not the camera; RGB uses the convolutional head, not the released attention head; symbolic analysis gives 3 cells on grayscale, 0 on RGB, unexplained.

Critical weaknesses

W1. Surrogate fidelity implying verification transfer is asserted, not measured. The abstract's 80% (world model) and 100% (GAN) sit on different surrogates, so the comparison invites a misreading. Table 1's 30k / 136k / 231k models already give the axis: verify each and plot resolved cells against SSIM.

W2. Table 3 belongs in the main text. 120 cells with the convolutional head, 55 with the small attention head, 10 with the released 1.73M one, which is the more accurate head and decides almost nothing. The headline RGB result does not hold for the benchmark as shipped; say so in the abstract, or show one sound abstraction for attention.

W3. Symbolic analysis decides 3 cells of 10,000 and 0 of 10,000 while being one of four pillars. One diagnostic settles it: compare the symbolic bound against the forward bound on the same handful of cells, to separate compounding McCormick looseness from a misallocated 3,600 s budget.

W4. Backward analysis decides 3,774 of 6,536 grayscale cells and 4,216 of 4,677 RGB, yet neither Algorithm 1 nor Appendix D says which intermediate step it uses or how sensitive the count is to that choice. Disable backward and report what remains; Table 5 partitions outcomes, it ablates nothing.

W5. Comparing 1,043 with 374 GPU-hours invites the wrong conclusion. Cost per resolved cell, or cost to reach a fixed verified fraction, is the number to report. The 3,600-hour RGB estimate is an extrapolation and reads as measured.

W6. Figure 5 has no legend and cannot be read quantitatively; RESOLVE hides whether symbolic analysis sees the refined subproblems; \mathbf{E} = {0}^2 in both case studies leaves the disturbance set dead notation; and code at submission rather than acceptance is what makes these bounds checkable.

Questions for authors

On W1 (fidelity → verification transfer): Can you report verified-cell counts for the three world-model sizes on Aircraft Taxiing (or a 10 \times 10 block of AEBS) alongside their RMSE/SSIM? Even a monotone trend on a small block would substantiate the central premise much more than the current argument-by-plausibility.

On W2 (attention head): Is the 10-cell result for the released attention head a soundness limitation (the abstraction is too loose to verify) or a budget limitation (10 GPU-hours is simply too few)? If you doubled the budget to 100 hours, would the count grow linearly (budget-limited) or saturate (abstraction-limited)? This distinction determines whether the open problem is engineering or fundamental.

On W3 (symbolic analysis): On 10 cells where symbolic analysis fails, what is the ratio of the symbolic bound to the forward bound at the same step? If it is > 10^2, the McCormick relaxation of the FiLM product over the unrolled horizon is the likely culprit and should be named; if it is O(1), the budget is the issue.

On W4 (backward analysis): What rule chooses the intermediate step at which the forward bound is handed to backward analysis, and how sensitive is the verified-cell count to that rule? A one-line specification in Algorithm 1 would suffice.

On W5 (cost): Can you report cost-per-verified-cell for both methods on the grayscale benchmark? The 1,043 vs. 374 comparison is misleading without it, and the 10-GPU-hour figure for the baseline's 2,669 cells suggests your method is much cheaper per cell on the easy region and expensive only on the hard region — which is the interesting story.

On the latent box: The world model's latents are held fixed within a rollout and varied across rollouts (Appendix B). Does the verification analysis assume \mathbf{z} can vary independently at every step (as stated in Sec. 4, "\mathbf{z} ranges over \mathbb{Z} independently at every step")? If so, this is a larger set of behaviors than the training distribution covers, which is sound but conservative; if not, the formalization should say so. This matters for interpreting the 80% RGB result.

3 comments
AuthorsOpenReview ↗

Fidelity, cost, and reproducibility

We thank the reviewer for their careful and constructive review, and provide responses below.

Fidelity and verification. We agree that we assert rather than establish the relationship between surrogate fidelity and verification. This is a general limitation of surrogate-based verification, and formalizing that relationship is work that we are actively pursuing. We will add the experiment the reviewer suggests, and verify world models of different sizes on a fixed block of cells. We note that the trend may not be monotone. In our experiments so far, larger world models produce better images, but their bounds are not tighter. We will update the paper to present fidelity as necessary but not sufficient, and we will state in the abstract that the guarantee is for the surrogate.

The attention head. We are able to analyze the attention head, albeit at great budgetary cost. Every cell we attempted was certified within the 10 GPU-hour budget. The released-size attention head is about 50 times more expensive per cell than the convolutional head. The cells we attempted were the easier half of the block, so we expect the number of verified cells to grow sublinearly with budget. We will move Table 3 to the main text and name the convolutional head in the abstract.

Symbolic analysis. The limitations of the symbolic approach on this problem are now better understood. The controller is saturated on the cells we certify, so the control interval at the deciding step is already exact and no tighter relaxation can change the verdict. The remaining cells are cells where the control varies within the cell, and branching is required to make progress on them. On the RGB benchmark the symbolic budget (120 seconds per cell) was also too short for a single call to finish. We therefore believe that the issue is neither the McCormick relaxation nor the budget alone, but that there is little looseness to remove on this problem. We will add the bound-ratio diagnostic the reviewer suggests to confirm this.

Backward analysis. We carry the forward bound to an intermediate step, and then show that the resulting set cannot reach the avoid set within the remaining horizon using the plant's stopping dynamics. This is why backward analysis succeeds on cells where forward analysis accumulates too much conservatism over the horizon. We will state the rule for choosing the intermediate step in Algorithm 1, and we will add the ablation the reviewer suggests.

Cost. We agree that cost per verified cell is the right comparison. The 1043 hours are the sum of per-process wall-clock time with up to four processes sharing one GPU, so the time on the card is roughly 300 hours. The cells the baseline certifies take us under 10 hours, and the rest of the budget was spent on the cells it could not decide. We will report cost per verified cell, and label the 3600-hour RGB estimate as a projection.

Presentation and reproducibility. We will add a legend and per-analysis counts to Figure 5, describe the data flow between analyses in Section 5, spell out the "Same" entries in Table 4, and point to the numerical-soundness paragraph from the main text. We keep the disturbance set in the formalism for generality, and will note that both case studies set it to zero. We intend to release the code with the final version of the paper.

The latent box. Yes, the latent varies independently at every step during verification, while training holds it fixed within a rollout. The certified set therefore strictly contains the training behavior. We will say so in Section 4.

Thanks for answering these instead of rebutting them. Three of your replies actually move me.

The 1,043 GPU-hours being four wall-clock streams sharing one card, so roughly 300 hours on it, is a correction I should have caught myself. My cost objection was weaker than I put it, and reporting per verified cell is the right fix. The 3,600-hour RGB figure being labelled a projection is all I was after.

On the attention head, certifying every attempted cell inside the budget is a fair answer, with the caveat that you took the easier half and expect sublinear growth.

The latent-box clarification I should have asked for; fair. Score stays where it was.

We have uploaded a revised PDF with changes marked in blue. Appendix F.5 reports the fidelity sweep you asked for, four world-model widths verified on the same block with their RMSE, head error, and closed-loop error. Section 6.3 reports cost per verified cell and the GPU accounting behind the 1,043 hours. Section 5 and Algorithm 1 state the rule for the intermediate step (the last step forward analysis completed), and on this benchmark the dynamics alone bound the backward set. Appendix F.4 reports the ratio of the symbolic bound to the forward bound at the same step, which is 1.000 on the cells we measured, so the explanation is saturation rather than budget. Section 5 points to the numerical soundness paragraph, Section 3 notes that the disturbance set is zero in both case studies, and Section 4 states that z varies independently at every step during verification. Thank you for engaging with our responses, and for a review that improved the paper.

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

This paper proposes RAC-WAM, which adds training-time adversarial consistency to Fast-WAM-Joint. It builds separate latent perturbations for the robot, object, and background, mixes action and future-video velocity discrepancies, assigns a conserved RMS budget with a local sensitivity score, and adds an auxiliary future-mask loss. Masks, the mask head, and PGD are removed at inference. Against the same backbone, LIBERO-Plus rises from 72.2% to 82.8%, while in-distribution LIBERO stays at 98.3% versus 97.7%. The ablation attributes most of that 10.6-point gap to mask supervision and adversarial consistency without region coverage. The sensitivity rule gives a larger radius to the region with the larger gradient.

Critical strengths
  1. The authors cast vision-based neural feedback control as reach-avoid verification of a state-based system whose sensor is a generative model, and the specification, the surrogate, and the scope of the guarantee are stated in one formalism.

  2. The authors show that the released grayscale emergency-braking system, left partly unresolved by prior work, can be decided over the whole initial grid, with each cell attributed to the analysis that decided it.

  3. The authors restrict the generator to layers that standard bound propagation already handles, and the limitations section stays inside what the experiments support.

Critical weaknesses
  1. The authors’ motivation is inconsistent with their main result. The paper is framed around GAN surrogates being too large and too hard to verify, but the grayscale benchmark is fully decided with the released cGAN. The proposed world model is not part of that result, and no experiment fixes the perception head and the initial grid and shows that the new surrogate is what makes the system decidable.

  2. The authors overstate the model. The surrogate is a supervised image decoder on a known analytic plant; on the RGB system, its only latent is sun altitude, held fixed during training and redrawn independently at every control step during verification. The paper gives no ablation against a deterministic decoder or a noise latent, and no comparison with the deterministic world models and VAE surrogates it cites as alternatives.

  3. The authors’ verifier is a cascade of existing analyses, and the piece they adapt does not carry the result. Falsification, forward bounds, refinement, symbolic unrolling, and backward reachability are prior methods run in a fixed order. Symbolic analysis decides three grayscale cells and no RGB cells. The previously unresolved grayscale cells are decided by forward bounds, refinement, and backward analysis.

  4. The authors’ RGB claim does not apply to the benchmark they name. They replace the released SAGAN and the released attention head with a much smaller decoder and a less accurate convolutional head, then leave 1,822 of the remaining cells unrun. A partial decision on this modified system is not a resolution of the open RGB benchmark.

  5. The authors never connect reconstruction quality to the safety claim. The large AEBS similarity gaps compare their decoder on their own renders with released GANs trained on unpublished data; on the one shared split the gain is small, and that system is not verified. The certificate concerns only the surrogate, and the paper does not show that a more faithful generator changes which cells are safe.

Questions for authors
  1. The grayscale benchmark is fully decided with the released cGAN. Which result shows that the world model, rather than backward analysis and refinement, is what makes verification succeed?

  2. For RGB, sun altitude is fixed along each training rollout but redrawn independently at every control step during verification. Which environmental variation does the certificate actually cover?

  3. The RGB experiment replaces both the released SAGAN and the released attention head, and leaves 1,822 cells unrun. Why is a partial decision on this modified system a result for the RGB benchmark?

  4. With the perception head and the initial grid held fixed, how do the released GAN, a deterministic decoder, and the grounded-latent decoder compare in perception error and in the decided cells?

Symbolic analysis decides 3 grayscale cells and 0 RGB cells. Which of the reported optimizations decides any cell that forward bounds, refinement, and backward analysis leave unresolved?

2 comments
AuthorsOpenReview ↗

Clarifying claims, Comparisons, and Limitations

We thank the reviewer for their review, and provide responses below.

The role of the world model on the grayscale benchmark. We agree that the world model plays no role in the grayscale result, and we do not claim that it does. The grayscale benchmark demonstrates our verification procedure on the released cGAN. The world model's contribution is on the RGB benchmark, where we argue that the released SAGAN is the wrong object for surrogate verification. It is less accurate on real frames than our world model (the released heads are off by 2.2 - 2.8 meters, compared to 0.17 - 0.38 meters for ours), and it is hard to bound even using the most accurate verifier settings. We verify the cGAN to show that the verification procedure works regardless of the surrogate, and we argue that GAN-based surrogates will not scale to realistic scenarios.

Environmental coverage. The certificate covers sun altitudes between 43.5 and 46.5 degrees, in any sequence over the horizon. The training data holds the altitude fixed within each rollout, so the certificate covers a larger set of behaviors than the training data contains, and is sound with respect to the original training distribution. We chose lighting as the latent because it is the environmental variation with the largest effect on the image. We will state the coverage explicitly in Section 4, and we are working on extending it.

The RGB benchmark. We did not claim to resolve the open RGB benchmark. Our resolution claim is for the grayscale benchmark. We instead show that the SAGAN surrogate is the wrong object for verification, and that our world model surrogate more accurately represents the scenario. We retained the benchmark's plant, controller, and scene, and we argue that the resulting closed loop (with our world model in place of the SAGAN) better represents the vision-based scenario that the SAGAN was supposed to. We also note that our procedure resolved over 99% of the cells we have evaluated, and we expect to verify the remaining cells before the camera-ready deadline.

Comparing surrogates with the head and grid fixed. We will add a comparison between the released SAGAN, a deterministic decoder, and our world model with the perception head and the initial grid held fixed, and report perception error and decided cells for each. We note that the bounds on the released SAGAN under its own head are vacuous, so the comparison will use our head for all three.

Symbolic analysis. We note that all pieces of the verifier were adapted (not just symbolic analysis). Nonetheless, the limitations of the symbolic approach on this problem are now better understood. The cells we certify are cells where the controller saturates (full brake or full release) over the entire horizon, and no tighter relaxation can change the verdict on those cells. The remaining cells are cells where the control varies within the cell, and branching is required to make progress on them. We will update the text to reflect the role of symbolic analysis on this benchmark.

The scope of the model. The reviewer is correct that the surrogate is a decoder on a known plant, and this is by design. The plant is known and exact, and learning it would weaken the guarantee. We note that our world model is a strict superset of deterministic world models. We will elaborate on the relationship to VAE surrogates in the related work.

The verification procedure. We do not claim that the individual analyses are new. Our contribution is the formulation that lets state-based analyses apply to vision-based closed loops, together with the modifications that make the composed system tractable (abstractions for the attention operation, relaxations for the FiLM product, batched bound propagation, and forward-backward integration). The reviewer is correct that the previously unresolved grayscale cells are decided by forward bounds, refinement and backward analysis. This is due in part to the position of symbolic analysis in our procedure. The easier cells are decided by forward analysis and branching before symbolic analysis is ever attempted, and the cells that reach it are the hard ones, where the controller is saturated and a tighter relaxation does not help.

Reconstruction quality and the safety claim. We agree that the certificate concerns the surrogate, and that we have not shown that a more faithful generator changes which cells are safe. This is a general limitation of surrogate-based verification, and we will state it in the abstract.

We have uploaded a revised PDF with changes marked in blue. The abstract now states that our guarantees hold for the surrogate systems, and Section 6.3 states that the RGB result is for our world model in place of the released SAGAN, not the released benchmark. Appendix F.2 compares the SAGAN, a deterministic decoder, and our world model with the perception head and the grid fixed, in perception error and in decided cells. Appendix F.1 compares all three against recorded rollouts. The related work states the relationship to VAE surrogates and deterministic world models. Section 6.3 and Appendix F.4 give the reason symbolic analysis decides so few cells. Section 6.3 states the environmental coverage of the RGB result, sun altitudes from 43.5° to 46.5°, varying independently at every step. If you feel that these changes answer your questions, we would be grateful if you would consider updating your score. If anything remains unclear, we would be happy to provide additional details.

Reviewer_5oX5OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the formal verification of vision-based neural feedback systems and proposes a framework that combines stochastic world models with reachability-based verification. The authors introduce a perception surrogate that generates images from system states and physically grounded latent variables, aiming to capture environmental variations while remaining amenable to formal verification. The world model adopts a deconvolutional architecture with FiLM-based latent modulation and is trained using simulated camera images. On the verification side, the authors develop a multi-stage procedure combining falsification, forward analysis, adaptive refinement, symbolic analysis, and backward analysis. Experiments on aircraft taxiing and automatic emergency braking benchmarks demonstrate improved image reconstruction fidelity and verification coverage compared with existing GAN-based surrogates. In particular, the proposed approach resolves all regions of the grayscale AEBS benchmark and over 80% of the RGB benchmark.

Critical strengths

-The paper addresses the formal verification of vision-based neural feedback systems, an important yet challenging problem in safety-critical autonomous systems. Integrating stochastic world models with formal verification is an interesting and potentially valuable research direction.

-The proposed world model incorporates physically grounded latent variables to represent environmental variations while maintaining a relatively compact and verifiable architecture. This design provides a reasonable approach to balancing perception fidelity and verification tractability.

-The experiments demonstrate improved image reconstruction fidelity and parameter efficiency compared with existing GAN-based surrogates. The verification results are also encouraging, particularly the complete resolution of the grayscale AEBS benchmark and the classification of over 80% of the RGB benchmark's initial state space.

Critical weaknesses

-The Introduction does not clearly establish the motivation for adopting stochastic world models over existing cGAN-based approaches. Although the authors discuss the limitations of cGANs, the specific advantages of the proposed world model and how they address these limitations are not sufficiently explained. Moreover, the first claimed contribution does not appear particularly novel, as related ideas have already been explored in prior work, such as Katz et al. (2022). The third contribution mainly summarizes experimental results rather than highlighting a methodological advance, and substantially overlaps with the Abstract. The authors should better articulate the technical novelty and distinguish their contributions from existing work.

-The Related Work section is somewhat fragmented, and several discussions are only loosely connected to the central research problem. A more focused organization around two main directions, namely stochastic world models and formal verification of vision-based control systems, would improve the clarity and relevance of the literature review. Other related techniques could be incorporated into these discussions where appropriate.

-The mathematical formulation lacks the rigor and precision typically expected in formal verification research. Several definitions and notational conventions are not presented in a sufficiently formal or mathematically precise manner, making the problem formulation difficult to interpret unambiguously. In particular, despite the emphasis on stochastic world models, the underlying randomness and its mathematical characterization are not clearly defined. Meanwhile, the paper devotes considerable space to illustrating the standard reach-avoid property, which is already well established in the field, while providing insufficient background on world models and stochastic modeling. The authors should adopt more rigorous mathematical definitions and focus the preliminaries on concepts directly relevant to the proposed approach. Additionally, the discussion in Lines 181–183 should better clarify the differences between the proposed method and existing work.

-The main methodology is described rather briefly, and the presentation does not adequately distinguish the proposed techniques from existing approaches. Several components are introduced primarily by referring to their similarities with previous methods, without clearly explaining what has been modified or improved. In particular, the discussion in Lines 210–213 requires further technical elaboration. The SSIM-based training objective should also be explained more clearly for readers unfamiliar with this metric. More importantly, Figure 3 could be used to systematically explain the overall pipeline, the roles of individual components, and their interactions, rather than focusing predominantly on implementation and training details. Figure 3 also appears not to be explicitly referenced in the main text.

-The proposed verification framework largely combines conventional techniques, including forward reachability analysis, symbolic analysis, backward analysis, and adaptive refinement. While their integration may be practically useful, it is unclear what fundamentally new verification techniques are introduced. For instance, the optimization mentioned around Line 296 is not sufficiently specified, making it difficult to assess its technical contribution. Algorithm 1 is also highly abstract and appears to primarily organize existing verification procedures without clearly identifying the proposed improvements. The authors should provide a more detailed explanation of the new algorithmic components and clarify how they differ from existing approaches and contribute to the reported verification improvements.

-The experimental results are promising, particularly the improved reconstruction fidelity of the proposed world model and the increased verification coverage on the AEBS benchmarks. However, several aspects of the experimental evaluation need further clarification.

i. The paper uses RMSE and SSIM on held-out images to demonstrate the fidelity of the proposed world model. While these metrics indicate improved image reconstruction quality, they do not necessarily imply more accurate closed-loop behavior or stronger safety guarantees. In particular, even small image reconstruction errors may lead to significant differences in control actions, especially near safety-critical decision boundaries. It would be helpful to provide additional evaluations of perception errors, control-action discrepancies, or closed-loop trajectory deviations under the generated observations. Such experiments would establish a stronger connection between image reconstruction fidelity and the practical relevance of the resulting verification guarantees.

ii. The proposed verification procedure combines several existing techniques, including forward analysis, adaptive refinement, symbolic analysis, and backward analysis. However, the individual contributions of these components are not sufficiently isolated. Although the paper reports the number of regions resolved by different analyses, it remains unclear how much additional benefit each component provides in terms of verification coverage and computational cost. In particular, the contribution of backward analysis beyond forward and symbolic analysis is not sufficiently demonstrated. A more comprehensive ablation study, including comparisons of different combinations and execution orders, would help establish the effectiveness of the proposed verification procedure.

iii. Although the proposed method completely resolves the grayscale AEBS benchmark, it requires 1,043 GPU-hours, compared with 374 GPU-hours reported by the baseline. Notably, approximately 1,030 GPU-hours are spent on regions that the baseline cannot resolve. This substantial computational overhead raises concerns about the scalability and practical applicability of the approach. A more detailed breakdown of the runtime of individual verification components, together with an analysis of how the computational cost scales with the state dimension and verification horizon, would help better assess the practical benefits of the proposed method.

iv. The RGB AEBS results are encouraging, but the reported verification coverage is obtained using a convolutional perception head rather than the original attention-based architecture. Although Table 3 provides additional results for attention-based perception heads, the released-size attention model verifies only 10 cells within the same computational budget, compared with 120 cells for the convolutional model. Consequently, the reported over 80% coverage does not directly demonstrate the verification capability on the original RGB benchmark. The authors should clarify the implications of replacing the original perception architecture and further discuss the limitations of the proposed approach when handling more complex attention-based perception networks.

Questions for authors
  1. The proposed framework appears to establish set-based reach-avoid guarantees rather than probabilistic safety guarantees. Could the authors clarify how randomness is modeled and whether the verification results can be transferred to the original vision-based closed-loop system?

  2. What is the main technical novelty of the proposed world model compared with existing conditional generative models? What advantages do physically grounded latent variables provide beyond improved reconstruction fidelity?

  3. Which components of the verification framework introduce new techniques rather than combining existing methods? In particular, what specific optimizations are introduced in verification procedure?

  4. Could the authors provide an ablation study to clarify how the world model and individual verification components contribute to the reported improvements in verification coverage and computational efficiency?

2 comments
AuthorsOpenReview ↗

Precision, Novelty, Fidelity, and Scaling

We thank the reviewer for their detailed review, and provide responses below.

Randomness. We agree that the formulation should be more precise. The observation model is a map g(s,z)g(s, z) from the state ss and a latent $z \in Z$, where ZZ is a box, to an image. The closed loop is defined for every choice of zz at every step, and the certificate holds for every sequence (z1,…,zT)∈ZT(z_1, \ldots, z_T) \in Z^T. The guarantee is therefore worst-case over ZZ rather than probabilistic. This is stronger than a guarantee over any distribution supported on ZZ. We call the model stochastic because it produces a set of observations for each state, and we will define it as a set-valued observation model in Section 3.

Novelty of the world model. Katz et al. (2022) also verify over a box of latents, but a cGAN latent box is not necessarily correlated with operationally relevant scenarios, while the deterministic world models of Geng et al. do not model observational variation. By design, our latents range over an interval grounded in a physical quantity, and our certificate is a semantically meaningful statement over possible operating conditions. Crucially, the resulting surrogate model is better able to represent the scenarios we wish to verify, while remaining more tractable for verification compared to GAN-based models.

Novelty of the verifier. We do not claim that the analyses are new. Our contributions are the formulation that lets us apply existing state-based verification algorithms to vision-based problems, and a series of optimizations that scale state-based verification tools. These optimizations include: new abstractions for the attention operation, relaxations for the FiLM product, new batching procedures for bound propagation, and forward-backward integration for verifying vision-based closed loop systems. We will update the paper to include a more thorough discussion of these optimizations.

Ablations. Tables 5 and 6 provide additional experiments to illustrate what each verification component contributes to the overall problem. Our verification procedure optimizes for efficiency using a cost-driven escalation ladder. It starts with cheap evaluation tools, and only attempts expensive evaluations when easier approaches have failed. We will conduct additional ablations to characterize the coverage and efficiency of our algorithm. We will report results as soon as we have them.

Fidelity and closed-loop behavior. We agree that image metrics alone do not guarantee closed-loop accuracy. We have compared closed-loop rollouts of our world model against rollouts of the real CARLA camera from the same 60 initial states. The distance traces differ by at most 0.09 m on average, and every rollout pair ends in the same outcome. We will add these results to the paper.

Computational cost. The 1043 hours are per-process wall-clock time summed over up to four processes sharing one GPU, so the time on the card is roughly 300 hours. The 2669 cells the baseline certifies take us under 10 hours. We note that cost scales with the horizon, and that the baseline verifies most of the easy cells. We will report a more detailed runtime breakdown to help understand how well the algorithm scales.

The attention head. We agree that the result does not demonstrate verification of the released attention head, and we will say so in the abstract. We note that the released attention head is significantly less accurate than our trained head (its mean error is 2.84 m compared to 0.17 m with our head). Verifying the released head will therefore provide a certificate for a model that does not represent the scenarios we actually care about. We also note that our algorithm is able to verify attention heads albeit at greater cost. We report a comparison of our algorithm on attention and convolutional heads in Table 3.

Presentation. We will reorganize the related work, shorten the reach-avoid exposition, explain SSIM, and reference/expand Figure 3 as an overview of the pipeline.

We have uploaded a revised PDF with changes marked in blue. Section 3 now defines the surrogate as a set-valued observation model and states that the guarantee is a worst case over the latent box, not a probability. The related work is reorganized around verification of vision-based neural feedback systems and stochastic world models, and states how our latents differ from a cGAN’s and from a deterministic world model’s. Figure 3 is referenced and explained in Section 4, and SSIM is explained in Section 6.2. Appendix F.1 compares closed-loop simulations with the surrogate against recorded CARLA rollouts (within 0.1 m, same outcome in 60 of 60). Appendix F.3 reports the ablations. Section 6.3 reports cost per verified cell (0.16 GPU-hours against 0.14 for the baseline) and the GPU accounting behind the 1,043 hours. The abstract and Section 6.3 state that the RGB result uses our convolutional head, not the released attention head. If you feel that these changes answer your questions, we would be grateful if you would consider updating your score. If anything remains unclear, we would be happy to provide additional responses

Reviewer_twq6OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies using stochastic world models instead of GANs for perception in vision-based feedback systems which are controlled by a neural network. Specifically, it argues that stochastic world models are better “perception surrogates” than GANs when verifying safety of the system is important.

Critical strengths
  • The formalization of neural feedback systems and explanation of how vision-based systems are compatible with the state-based formalism seems interesting, although many 14 elements of the tuple are not clearly defined.
  • The formalization of a “safe” system is also good (Equations 2 and 3) and novel
  • The empirical results seem promising and point to an advantage of stochastic world models over GANs
Critical weaknesses
  • The authors claim that their formalization of vision-based neural feedback systems allow techniques for state-based systems to apply to the vision-based setting. However, it is unclear what fundamentally differentiates analyzing this (inexact) observation model from standard state-based techniques beyond coarsening/bounding the observation space. It seems like the reachability techniques being applied are conceptually identical to existing neural feedback verification methods.
  • The authors claim one of their main contributions is a verification procedure that combines “falsification, adaptive refinement, symbolic, and backward analysis” (L70). However, all of these subroutines are inherited from existing work on state-based neural feedback systems, as noted in the related work section.
  • The paper seems incomplete, i.e., Table 2 notes that 1,822 cells of the world model had not yet been analyzed. As the limitations note, “no formal guarantee relates the surrogate to the camera it replaces” , and the benchmark with the released attention head has not yet been verified and points out a confusing observation then says “we have not yet determined why” with next steps that read like natural results that should have been included in the paper.
Questions for authors
  • The distinction between exact state-based systems vs. observation-based ones is not clear. This sentence in particular is confusing: “Recent work verifies state-based neural feedback systems, whose controllers compute actions directly from the state, but the vision-based case remains largely open” (L36-38). Is the only difference that observations must be coarsened in order for the system to represent and process them?
  • To what extent does this idea extend beyond vision-based cases to any case where there are high-dimensional /noisy observations? Nothing about the formalization seems specific to vision.
  • The paper notes that the model only uses layers which standard verifies bound - how restrictive is this?
2 comments
AuthorsOpenReview ↗

Generality + Incompleteness

We thank the reviewer for their review, and provide responses below.

Distinguishing state- and vision-based systems. We agree that state- and vision-based neural feedback systems are variants of the same underlying problem. However, previous research does not treat them in this manner. This is due in part to the difficulty of verifying vision-based problems, and in part because algorithms for state-based problems do not scale to vision-based problems. Vision-based problems do not simply coarsen the observation space, they increase the dimensionality of the problem by several orders of magnitude, and better reflect verification scenarios. Our contribution is an attempt to demonstrate this fundamental similarity by showing that, using the right abstractions, we can verify vision-based neural feedback systems with tools built for state-based ones. We achieve this by reformulating the vision-based problem, and by scaling state-based algorithms. We recognize that the paper may not state this clearly, and will rewrite it to better represent our contributions.

Generality beyond vision. We agree that nothing in the formalism is specific to cameras, and we are actively exploring demonstrations beyond images. Any high-dimensional observation can be modeled in this manner, and we are excited about the prospects of using world models for verifying such problems. Our experiments are confined to vision because vision represents the primary domain where these ideas are being explored. We will update the paper to reflect the broader scope of our contributions.

Architectural restrictions. These restrictions reflect the limitations of state-of-the-art verification tools. Nonetheless we are able to support a wide range of layers, including DCGAN-style decoders. We are also able to verify attention-style layers (albeit at greater cost, as seen in Table 3). We will move the analysis of Table 3 to the main text to highlight the tradeoffs we made.

Incompleteness. We agree that the RGB results are incomplete, and we are actively working on verifying the rest of the benchmark. We have verified over 80% of the RGB problem, compared to 0% by the previous state of the art. Crucially, our procedure has verified over 99% of the cells it has attempted, and we expect to verify the problem before the camera-ready deadline. The limitations of the symbolic approach on this problem are now better understood. The controller is saturated on those problems, and branching is required to obtain any additional insights. We will update the text to better reflect the status of the paper.

We have uploaded a revised PDF with changes marked in blue. The introduction now states that nothing in the formulation is specific to cameras. The contributions say plainly that the analyses come from work on state-based systems and list what we add (abstractions for attention and the FiLM product, batched bound propagation, and the forward-backward combination). Section 5 and Appendix F.3 describe the relationship between forward and backward analysis, with ablations included. The status of the RGB benchmark is stated in the abstract and Section 6.3. If you feel that these changes answer your questions, we would be grateful if you would consider updating your score. If anything remains unclear, we would be happy to respond as soon as we can.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to reviewers

We thank the reviewers for their attention to our work. There are a few recurring points, and we attempt a response in this section.

  1. The verifier's contribution. We do not claim that the components of the verification algorithm are novel, in fact we rely on existing algorithms to make our central claim. Our contribution is that our formulation allows us to apply algorithms that were previously built for state-based neural feedback systems to vision problems. Our formulation, combined with modifications to, and optimizations of existing algorithms makes the vision-based benchmarks tractable. Previous approaches to solving those problems did not scale (as evinced by the unresolved cGAN and SAGAN benchmarks). We appreciate the reviewers highlighting the rhetorical failures in our presentation. We will update our framing to better reflect our contributions.

  2. The contributions of backward analysis. Backward analysis is used to solve 3774/10000 cells of the cGAN benchmark. This includes 1105 cells the baseline could not certify using forward-only/symbolic analysis. We will update the paper to clarify the relationship between forward and backward analyses.

  3. The RGB benchmark. We argue that the SAGAN benchmark is the wrong object for surrogate verification. It is less accurate (often off by 2.2 - 2.8 meters) on real frames, and difficult to certify (often yielding vacuous bounds on the most accurate solver settings). A world model reproduces the scene more accurately (errors between 0.17 and 0.38 meters), and can be verified using existing frameworks. The underlying benchmark remains the same. We keep the plant, controller, and scene being tested in the original paper. We note that the purpose of the original benchmark was to describe methods to verify vision-based controllers, and our approach provides a way to do so in a more realistic manner. We will update the paper to reflect this point.

  4. Guarantees about the surrogate. The verification certificate is for the surrogate model. This is a general limitation of surrogate-based verification frameworks. We agree that establishing a formal relationship between the sensor and the surrogate is important, and this is work we are actively pursuing. At the moment, the best metric we have is structural similarity, and empirical comparisons of real vs surrogate trajectories.

Dear Reviewers,

Thank you for your thoughtful reviews. The authors have provided detailed responses to the concerns raised.

I would appreciate it if you could revisit the author responses and clarify whether your main concerns have been adequately addressed, particularly which issues remain critical to your overall assessment.

Given the differences in the initial recommendations, I also encourage you to consider and discuss one another's key arguments to help clarify the remaining disagreements and develop a more consistent understanding of the paper's merits and limitations.

Please feel free to update your ratings if your assessments have changed following the discussion.

Thank you for your continued contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.