acceptodds
Withdrawn from ICLR 2027

Does Calibrated Surprise Detect Physical Violations? Auditing Conformal Monitors for Latent World Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Prediction error in a latent world model, or surprise, can flag an unexpected transition. What makes that alarm evidence of a physical inconsistency? We introduce VoE-Bench, a violation-of-expectation benchmark combining an audit of 29 checkpoints in three environments with focused tests of reference sensitivity, physical selectivity, and decision utility. In TwoRoom, changing the clean (unedited, default-rendering) reference moves alarms on the same 1,000 benignly recoloured episodes from to , although both nominal audits meet tolerance. Fresh references (9,600 episodes) confirm this sensitivity and reveal discrepancies from a point forecast frozen before scoring. In PushT, an action-aware multilayer perceptron (MLP) detects more action mismatches than its action-free counterpart under separate calibration. At approximately empirical nominal false-positive rate, the original head has no resolved advantage over the action-aware MLP sharing its encoder and inputs in three-frame onset-window detection ( percentage points; interval ). Under constant recolouring, mean teleport ordering above an appearance-only twin is when the teleported object is recoloured and when the other object is (other-minus-same difference ; interval ). TwoRoom readouts recover position despite unselective residuals, and later PushT residuals reveal action evidence masked by an appearance transient. Separate planning tests resolve no success gain and expose the cost of mismatched screening references. The audit distinguishes reference sensitivity, appearance-dependent ordering, and temporal masking, identifying what must be checked before interpreting or acting on an alarm.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
latent world models, joint-embedding predictive architectures (JEPA), conformal prediction, uncertainty quantification, violation-of-expectation, runtime monitoring, physical consistency, distribution shift

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

This paper introduces VOE-Bench, an evaluation framework for auditing whether conformally calibrated latent world model prediction errors can differentiate physical inconsistencies from benign changes. The authors experimented across 3 environments, and show that sensitivity to calibration references, appearance-dependent detection, and find no significant improvement in downstream planning performance.

Critical strengths
    1. The paper performs a rigorous set of experimental controls (nominal calibration, physical selectivity, decision utility), strengthening the credibility of empirical findings.
    1. The paper provides strong evidence of calibration sensitivity and object dependent effects, with quantitative analyses and uncertainty estimates. These findings can be useful to understand limitations of interpreting latent prediction errors as indicators of physical violations.
Critical weaknesses
    1. Despite being among the central contributions, technical details and justification for VOE bench is severely limited in the main text (most are buried deep in the appendix). Furthermore, the environments and several interventions have been explored in prior works, making it challenging the contribution and novelty of the benchmark and its analyses.
    1. While the authors audit 29 checkpoints, key findings rely on a small number of selected models and environments. The proposed diagnostics do not translate into validated online monitoring methods, and downstream planning experiments demonstrate no statistically significant improvements.
Questions for authors

-1. Could the authors clarify which aspects of the benchmarks and its empirical findings constitute genuinely new contributions beyond existing violation-of-expectation evaluations? A more explicit comparison would help assess the significance and novelty of the contribution.

  • 2. To what extent do the reported phenomena generalize across the other model families and environments? This is important to assess whether the paper establishes general limitations of latent surprise.
    1. The experiments primarily use the upstream PushT checkpoint, which exhibits relatively weak action-mismatch detection. Could the authors clarify how this choice affects the interpretation of the negative planning results.
    1. The paper defines timely detection using the interval [j, j+2], while acknowledging that an action mismatch cannot affect the observed position until j+1. Could the authors clarify how many reported timely action-mismatch alarms occur at j, before the intervention becomes observable? Would excluding such alarms materially change the primary predictor comparison?
    1. IN Section 2.2, it appears that PushT contributed windows to world-model training, so does this mean that the evaluation is not independent of model training? If so, this is a potential confound.
    1. The description for VOE bench and its itergration needs to be incorporated to the main paper, with much clearer explanation for readers to digest.
Reviewer_acs1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether calibrated latent prediction error, called “surprise”, can reliably indicate physical violations in latent world models. It introduces VoE-Bench, an audit framework covering 29 checkpoints across TwoRoom, PushT, and Cube. The framework separates three questions: nominal calibration, physical selectivity, and decision utility.

The main empirical finding is that nominal conformal calibration does not make an alarm physically meaningful. In TwoRoom, two clean reference sets produce recolouring alarm rates of 0.5% and 82.4% on the same benign episodes. In PushT, detection depends strongly on the prediction head, operating point, recoloured object, and temporal aggregation. The paper also shows that useful state information can remain accessible in the latent representation even when the residual is not physically selective. Finally, the tested replanning and subgoal-screening rules provide no resolved improvement in task success.

The validation is extensive. It includes matched physical and appearance-only interventions, independent nominal audits, paired bootstrap analyses, common-FPR comparisons, state readouts, temporal diagnostics, and controlled planning experiments. The paper also documents protocol deviations and post hoc corrections.

Critical strengths

The paper identifies a concrete failure in interpreting calibrated surprise: The TwoRoom result is strong and easy to understand: changing only the clean calibration reference changes the benign recolouring alarm rate from 0.5% to 82.4% on the same queries. Both original nominal audits satisfy the paper’s tolerance. This directly shows that nominal calibration does not establish robustness to benign appearance changes. This result is consequential because it challenges a common interpretation of calibrated prediction-error monitors without claiming that conformal calibration itself is invalid.

The evaluation separates detection, selectivity, and downstream utility: The paper does not treat a high intervention-detection rate as sufficient evidence. It uses appearance-matched twins, paired ordering, temporal diagnostics, and planning tests. This exposes cases where appearance dominates the alarm, physical evidence is temporally masked, or acting on an alarm does not improve outcomes. This separation is a useful methodological contribution. It gives future papers a more demanding standard for claiming that a world-model residual detects physical inconsistency.

Critical weaknesses

The strongest conclusions rely on a narrow and partly training-exposed evaluation setting: The primary predictor and object-appearance conclusions come mainly from one innovation-trained PushT checkpoint. Every PushT source episode contributed windows during world-model training. The reported uncertainty is conditional on the selected checkpoint and does not cover variation across independently trained world models. This matters because the paper presents VoE-Bench as a general audit methodology, but the clearest findings may partly reflect one representation, checkpoint, data pool, or training seed. The broader benchmark shows substantial variation between checkpoints, including a collapsed TwoRoom seed. Stronger evidence from unseen episodes or independent model-training runs would materially increase confidence in the general conclusions.

Questions for authors

Clarification requests

  1. What is the intended output of VoE-Bench for a new monitor? Is there a fixed set of mandatory tests and reported metrics, or must users select diagnostics based on observed failures? A clear answer would affect whether I view VoE-Bench as a reproducible benchmark or mainly as an audit methodology.

  2. Which main conclusions do the authors expect to generalise beyond the tested checkpoints? Please separate claims about the audit methodology from empirical claims about latent world models. In particular, clarify which claims are supported across multiple independently trained checkpoints and which are supported only by the primary PushT checkpoint.

Reviewer_byq1OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a benchmark for isolating 'surprise' in world simulators. In particular, the goal is to distinguish substantial physical violations from superficial issues (e.g. re-coloring of appearance). The paper proposes to do this via a conformal inference procedure, using the distance between the predicted and true latent states (sts_t) as the conformity score. While detection is possible, the results are somewhat cautionary as the benchmark passes the audit under shift and surprise-based planning seems ineffective.

Critical strengths

Testing / auditing world simulators for their fidelity is an important problem.

Critical weaknesses

critical weakness: I don't understand how the paper claims to propose a benchmark: if one is proposing a benchmark, then various methods should be tested---in this case, for detecting the physical vs superficial anomalies. Instead, only one conformal inference procedure is selected---one that doesn't even seem to be standard or well-established---and many other reasonable alternatives are ignored. In particular, a well-validated benchmark should test against sequential change point detection, uncertainty-based monitors (e.g. ensemble disagreement), latent density scores, learned detectors like CheckVLA, and conformal variants with temporal / shift awareness.

major weakness: the writing is extremely dense and hard to follow. moreover, little-to-no background is given or any strong motivation for the approach.

Questions for authors

My main request is to justify why conformal inference should be the only method considered, as opposed to others with much longer history and IMO seem more natural for the problem, such as sequential change point detection.

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

Related papers

Discussion (0)

Sign in to comment.