acceptodds
Under review as a conference paper at ICLR 2027

Snotra: Continually Adapting Irrigation World Models under Physical Change

Abstract

Physical agents must revise imperfect world models as the systems they control change. However, persistent prediction error alone does not reveal which change is identifiable, which revision deserves belief, or whether that revision matters for action. We study this problem in precision irrigation, where partially observed soil–crop–weather dynamics interact with constrained, long-horizon interventions. We present Snotra, an evidence-governed framework that separates world-model revision into three decisions: KNOW, whether competing explanations are distinguishable from available evidence; BELIEVE, whether a candidate survives validation, falsification, physical-consistency, safety, and provenance checks; and ACT, whether an admitted model changes a legal decision and improves an independently evaluated physical outcome. We further introduce SnotraBench, a stage-resolved evaluation protocol for world-model revision under physical change. On 35 matched governance units, Snotra accepts all 10/10 supported revisions while rejecting all 25/25 unsupported ones; Heldout-Only accepts the same 10/10 supported revisions but also 10/25 unsupported revisions. The paired reduction in unsupported admissions is 0.286 (95% cluster-bootstrap CI [0.143, 0.429]). In a one-check-removed replay, identifiability and physical-consistency are the only checks whose removal increases false admissions on this scenario set. Finally, a controlled drift sweep exposes a prediction-to-decision threshold: prediction discrepancy increases as physical drift grows from 0% to 50%, but the irrigation action remains unchanged through 30%; at 50%, revision changes the action by 10 mm and reduces independently evaluated cost by 110 protocol units. These results show that predictive repair, model admission, and decision value are distinct quantities and motivate evaluating adaptive world models by both admission selectivity and prediction-to-decision transmission.

Primary area:
learning on time series and dynamical systems
Keywords:
world models, physical agents, model adaptation, physical world-model revision, dynamical systems, system identification, identifiability, model-based control, physics-guided learning, decision-focused evaluation, precision irrigation, benchmarks

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

This paper introduces Snotra, a framework for updating physical world models when the environment changes. It separates model revision into three steps: KNOW (whether a change can be identified), BELIEVE (whether the updated model is supported by evidence), and ACT (whether the update improves decisions).

The authors evaluate the framework using irrigation data and controlled simulations. On 35 test cases, Snotra rejects all unsupported revisions while accepting all supported ones, outperforming a validation-only baseline in revision selectivity. The experiments also show that better predictions do not always lead to different irrigation decisions. The paper highlights the importance of checking model reliability and actual decision value separately.

Critical strengths
  1. A clear and useful approach to world-model adaptation. Separating model identification, validation, and decision-making is a meaningful idea. The framework addresses an important problem: improving prediction accuracy does not necessarily mean that a model update is trustworthy or useful for control.

  2. Careful evaluation of different stages. The authors compare Snotra with several baselines and examine which validation checks actually affect the results. I appreciate that the paper clearly reports its limitations rather than claiming that the framework works in every situation.

Critical weaknesses
  1. Limited experimental evidence for general adaptation. The main evaluation uses only 35 predefined test cases, and the decision-change experiment is based on a single controlled scenario. Although the results are encouraging, they do not yet demonstrate that the framework can handle a wider range of physical changes or achieve reliable continual adaptation in practice.

  2. Better model admission does not clearly improve control performance. Snotra rejects more unsupported revisions than the validation-only baseline, but its average irrigation cost is actually higher. This raises an important question about the practical benefit of the additional checks. The paper also combines several existing ideas, so the methodological novelty beyond their integration could be clearer.

Questions for authors

Q1. Snotra rejects more unsupported revisions, but its average control cost is higher than Heldout-Only. How should this trade-off be interpreted, and under what conditions would Snotra provide better practical outcomes?

Q2. The evaluation relies on a relatively small set of predefined scenarios, and only two admission checks affect the results. What evidence supports the usefulness of the full framework beyond these specific cases?

Reviewer_6Sy6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Snotra, a framework for analysing deployed "world models" under changing environments, using an irrigation example as a case study. It separates three different components for analysis, notably "KNOW" (identification of distinguishable explanations), "BELIEVE" (About admitting candidates through evidence checks) and "ACT" (about measuring effects on actions and independently evaluated outcomes. This framework uses "gray-box" soil-water models, and the paper proposes a benchmark named Snotrabench.

The validation is done through 7 configurations including overlapping subsets and counterfactual perturbations. The paper studies 35 governance scenarios, of which Snotra accepts 10 supported revisions and reject all the remainder 25 unsupported revision, with heldout-ounly accepting unsupprted revisions. The paper includes additional tests for the KNOW component of the framwork, and additional results show lower mean absolute error across several episodes.

Critical strengths

Snotra's decomposition of world model analyses into knowability, admission and action is interesting and provides a useful evaluation framework. This explicit separation allows for better disentangling better predictions from actual evidence of successful control or diagnosis. Further, the matched heldout only comparison and the reported negative results are useful for fully understanding the scope and limits of the contributions. This level of candor I think is positive to see on a paper and I evaluate it accordingly.

Critical weaknesses
  • I found the paper to be very hard to read in many areas, and I believe it sholuld benefit from a rewrite, prioritising conciseness and clarity. In this regard, I found that the paper (and Figure 1 in particular) tends to mix the more general-purpose framework from its irrigation application, making it hard to separate them.
  • Relatedly, the central result depends on a scenario that is purely constructed by the authors themselves, which is also under-described and therefore difficult to reproduce. It is very hard to grasp how the evaluation cases are constructed, and whether or not they are realistic. There are no external benchmarks or baselines, making the framework, application and validation to be too self-contained, unspecified and difficult to reproduce.
  • In this regard, the reproducibility statement refers to “accompanying anonymous supplementary source,” but I could not locate it. Please identify where it is available and clarify whether it includes the Snotra implementation, scenario generation, gate definitions and thresholds, and scripts to rerun experiments, or only scripts to regenerate figures and tables from existing outputs.
  • The evidence in this paper is insufficient to claim that the framework is useful beyond the proposed irrigation application. Using a single experimental setup on irrigation, which seems entirely handcrafted for purpose, is insufficient in my eyes for validating a general-purpose world model framework. Besides, no positive continual-learning benefit or overall control superiority is demonstrated, and comparisons beyond internal variants are missing.
Questions for authors
  • How were the 35 governance scenarios generated and labeled?
  • How is "harmful deployment" defined?
  • Will the authors release the full code?
  • How does this framework generalise to other world model applications?
Reviewer_nJveOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper is motivated by the important observation that "updating a model whenever prediction errors persists is not enough", and it proposes an auditable world-model revision lifecycle, Snotra, for closed-loop decision-making, separating KNOW (whether alternative explanations for observed model mismatch are distinguishable), BELIEVE (whether a candidate revision is sufficiently supported to be admitted), and ACT (whether an admitted revision changes a legal action and improves an independent physical outcome). The paper also introduces a benchmark that evaluates where evidence propagates or stops across the revision lifecycle. The framework is demonstrated in precision irrigation using multiple field-data configurations and controlled parameter, structural, and model-drift scenarios, in comparison with static, direct-update, and heldout-only baselines.

Critical strengths
  1. The KNOW–BELIEVE–ACT framework/revision lifecycle is thoughtfully designed and well-motivated, with a clear separation between determining what can be learned from a model mismatch, deciding whether a candidate revision should be admitted, and assessing whether an admitted revision actually matters for downstream decisions.

  2. The evaluation is carefully designed and covers a diverse set of revision and failure scenarios, including cases where evidence should stop at different stages of the lifecycle. The authors are also cautious in interpreting the empirical evidence and clearly distinguish what the experiments establish from broader capabilities that have not been demonstrated.

Critical weaknesses
  1. Although the proposed revision lifecycle is thoughtfully designed, I am not fully convinced about the added value of such a carefully crafted framework over a reasonably well-designed conventional model-revision process. In the relatively simple and constrained world-model setting considered in the paper, a practitioner might already use some combination of model diagnosis, independent validation, physical and safety checks, and downstream decision evaluation, even if these steps are not formalised as rigorously as in the proposed Snotra framework.

  2. The static, direct-update, and heldout-only baselines are useful for understanding the role of different safeguards, but they appear closer to ablations or simplified variants of the proposed lifecycle than to realistic alternative model-revision pipelines. Without comparison against a reasonably careful conventional revision process, it is difficult to determine how much practical benefit comes specifically from the proposed lifecycle.

Questions for authors

I understand the framework and find the overall design sound. My main reservation is that I am not yet convinced of its added value over carefully designed conventional model-revision practice, particularly given the baselines used in the experiments. Could you therefore explain:

  1. what additional value Snotra provides beyond a carefully designed conventional model-revision process?

  2. how the current baselines demonstrate this added value, given that they appear closer to simplified variants or ablations of the proposed lifecycle?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.