acceptodds
Under review as a conference paper at ICLR 2027

Relational Learning as World Modeling

Abstract

Relational learning typically predicts from a database state, but decision making asks a different question: how will that state change under an action? In operational systems, the answer depends on system-specific dynamics that may not be visible in the database itself. We view each such system as a relational "world" and ask whether a model can infer how it behaves from observed executions. We introduce TabWM (Tabular World Model) and WorldRel, a generator of executable relational worlds for pretraining. Given executions from an unseen world, TabWM uses them as context to predict the sparse structured edits caused by a new action without parameter updates. Across 512 held-out worlds, complete-edit F1 rises from 0.016 with no observed executions to 0.685 with 64, where a correct edit must match its operation, relational location, and value. Because these predictions are explicit relational edits, they can be applied recursively to imagine candidate futures and compare actions before they are taken. Broader search over these predicted futures improves the realized return of selected action sequences. Together, these results suggest that relational dynamics can be learned in context, broadening relational learning from modeling system state to modeling how that state changes under action.

Primary area:
foundation or frontier models, including LLMs
Keywords:
foundation models, relational deep learning, relational data, transformer, world models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_L9DeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

This paper investigates action-conditioned state transition modeling in relational databases. While standard relational deep learning focuses on static target prediction (S→yS \to y), this work formulates relational database updates as dynamic transitions where an action AA on state SS produces a sparse set of structured edits ΔS\Delta S (INSERT, UPDATE, DELETE across multiple tables). Adopting the Prior-Data Fitted Network (PFN) paradigm, the authors introduce WORLDREL, a procedural synthetic data generator for relational schemas, actions, and transition rules. They propose TabWM, which combines a Relational Transformer backbone with a DETR-style set prediction decoder and gated affine/additive numeric transfer heads to infer system dynamics from KK demonstrated transitions in context without parameter updates. Finally, TabWM is evaluated on synthetic held-out worlds and tested as a recursive rollout model for 3-step candidate action sequence ranking.

Critical strengths
  1. An Interesting Attempt to Extend Relational Learning to Action-Conditioned Dynamics: Moving relational learning beyond passive, static property prediction into modeling state transitions under external actions is a sensible and exploratory conceptual direction. It provides an interesting angle for exploring how tabular representations might interact with dynamic decision-making environments.
  2. Plausible Architectural Adaptation for Discrete Database Edits: Formulating multi-table database changes as an unordered set of sparse edits via DETR-style bipartite matching is a reasonable engineering choice for capturing discrete database operations. The inclusion of action/consequence shuffling experiments also provides an intuitive sanity check that the model is sensitive to behavioral correspondence in context.
Critical weaknesses
  1. Unclear Positioning and Baseline Representativeness Relative to TabPFN: The underlying learning mechanism of TabWM shares the Prior-Data Fitted Network (PFN) paradigm (Müller et al., 2022; Hollmann et al., Nature 2025). However, the comparison with TabPFN would benefit from greater clarity and fairness. In Table 1, "Flat-TabPFN" is evaluated on fixed-width vectors without schema names or relational edge topology, which naturally limits its performance. Given that the authors cite the recent relational foundation model TabPFN-Rel (2026), it remains unclear why a multi-table PFN extension equipped with transition heads was not considered as a baseline, and what unique structural necessity TabWM offers beyond a natural adaptation of the PFN framework.

  2. Practical Motivation of the Setting and Unjustified Planning Horizon (H=3):

  • Conceptual Tension in Setting: We acknowledge that real-world operational database benchmarks with aligned before-and-after action traces currently do not exist in the public domain, which reasonably justifies the use of a procedural synthetic environment (WORLDREL). However, there remains an inherent conceptual tension in the problem motivation: in real-world enterprise databases, business logic and cascade updates are typically deterministic and hardcoded into backend software services. If the system rules are already specified in code, the practical necessity of training an approximate neural world model to predict state transitions remains somewhat ungrounded.
  • Unmotivated Horizon Choice: In Section 8 and Appendix B.4, the model-based planning is evaluated exclusively at a fixed horizon of H=3H=3. The paper does not provide domain statistics, workflow analysis, or empirical justification for why 3 steps was selected as the representative decision depth.
Questions for authors

I will consider raising my score if the authors can address my concerns below:

  1. On TabPFN Positioning and Baseline Comparison: Could the authors clarify the conceptual and empirical boundary with TabPFN? Specifically:
  • (a) How was Flat-TabPFN constructed, and does stripping relational topology and schema context account for its low score?
  • (b) Could a relational variant of the PFN paradigm (such as adapting TabPFN-Rel with edit-prediction heads) serve as a more representative baseline, and what distinct capabilities does TabWM provide over such an extension?
  1. On Evaluation Settings (Planning Horizons and Real-World Traces):
  • (a) Choice of H=3H=3 and Alternative Settings: Could the authors clarify the rationale behind fixing the planning horizon specifically at H=3H=3, and why experiments under alternative horizon settings (e.g., H=1,2,5H=1, 2, 5 or variable-length horizons) were not explored? Is this because representative decision workflows in these worlds naturally conclude within three steps, or was this guided by search complexity or compounding error considerations?
  • (b) Feasibility of Real-World Traces: Supplementing the evaluation with real-world operational execution traces (even on a small scale from enterprise change logs or workflow benchmarks) would be ideal. We fully recognize that aligned before-and-after database benchmarks are difficult to obtain in the public domain and understand if this is not possible for the rebuttal, but any preliminary evidence or discussion on real-world feasibility would be welcomed.
5 comments

Thank you for the thoughtful review and for recognizing the value of extending relational learning to action-conditioned dynamics and structured edit prediction! We would like to address your concerns below:

Q1(a) / W1. Flat-TabPFN and relational information.

Flat-TabPFN converts action-conditioned transition tasks into fixed-width feature-based prediction problems. As detailed in Appendix B.2, its features summarize action bindings, relational links, column statistics, and observed changes. It therefore retains relational and behavioral cues, although not the full relational topology or the structure of individual executions.

Does this information loss explain its lower score? Table 1 cannot isolate that effect, since the models also differ in training and prediction interfaces. We therefore use Flat-TabPFN as a feature-based reference, rather than a matched topology ablation. Our controlled experiments in §6 and Appendix F.5 separately examine behavioral correspondence and model components. We will clarify this distinction.

Q1(b) / W1. TabPFN-Rel and TabWM.

TabWM's contribution is a framework for identifying system-specific relational dynamics from executions in context, not simply a new PFN output head. The PFN paradigm learns from tasks sampled from a prior [1]. TabPFN-Rel [2] uses relational feature synthesis for entity-level prediction. In contrast, WORLDREL supplies a prior over executable dynamics (§4.1), TabWM represents aligned state–action–edit executions (§4.2), and its decoder predicts variable-size edits with relational pointers and numeric transfer (§4.3). These edits support recursive future-state prediction (Eq. 8).

The distinction matters empirically. In §3, identical states and actions yield different edits in 65% of paired-world probes, making observed behavior essential for identifying the active dynamics. In Figure 4, shuffling actions reduces Edit F1 from 0.664 to 0.041 at K=64; Table 2 shows adaptation to reassigned action-to-mechanism mappings; and the matched numeric-transfer ablation improves Edit F1 from 0.476 to 0.616 (Appendix F.5). A TabPFN-Rel-based transition model is possible, but adding an edit head alone would not supply the execution-conditioned training prior or representation. We will make this distinction clearer.

Q2(a) / W2. Planning horizon.

H=3 is a controlled test of composing predicted transitions, not a claim about typical enterprise workflow lengths. We additionally evaluated H=1, 2, 3, and 5 on 40 held-out worlds (K=16, N=8, shared candidate prefixes), with 95% world-bootstrap confidence intervals:

HorizonSpearman ρ [95% CI]Normalized regret ↓ [95% CI]
H=10.376 [0.239, 0.512]0.148 [0.068, 0.243]
H=20.450 [0.319, 0.577]0.226 [0.130, 0.329]
H=30.414 [0.277, 0.551]0.207 [0.119, 0.305]
H=50.486 [0.352, 0.613]0.205 [0.116, 0.306]

Ranking correlations remain positive with confidence intervals above zero at every horizon, showing that the ranking signal is not specific to H=3. This is a separate horizon-sensitivity test; §8 compares candidate selection methods at H=3. We will clarify the evaluation purpose.

Q2(b) / W2. Practical motivation and real-world traces.

A deterministic backend is not necessarily a simulator available to an external agent. An agent operating through an API may observe actions and database changes but lack access to the rule engine or a safe way to execute candidate actions in advance. Predicting consequences from past executions is useful in precisely this black-box setting. World of Workflows [3] independently studies forward audit prediction under hidden business rules, including cascading changes across tables.

We also test transfer to real database structures in §7, using topologies from THE JOIN and value distributions from RelBench. At K=64, TabWM obtains 0.657 Edit F1 on imported topologies versus 0.671 on matched synthetic ones. These are not real operational execution traces; such traces could be constructed by aligning application action logs with database audit records. We will clarify this distinction and the intended deployment setting.

Thanks again for the valuable feedback! We hope this can address your concern adequately. If you find the clarifications satisfactory, we would appreciate your consideration in re-evaluating the manuscript.

[1] Müller et al. "Transformers Can Do Bayesian Inference." ICLR (2022).
[2] Hayler et al. "Advancing Open and Reproducible Relational Learning: RelArena-α, TabPFN-Rel and RPI." arXiv:2608.16319 (2026).
[3] Gupta et al. "World of Workflows: A Benchmark for Bringing World Models to Enterprise Systems." arXiv:2601.22130 (2026).
Reviewer_L9DeOpenReview ↗

Thanks For Response

Thank you for the response and the additional experiments. The H = 1, 2, 3, 5 evaluation addresses my concern regarding the fixed horizon, and the clarifications on the external API-agent setting, Flat-TabPFN, and TabPFN-Rel improve the paper's positioning. Accordingly, I have raised Presentation and Soundness to 3.

However, I am maintaining Informativeness at 2 and my overall rating at 2 (Weak Rejection) due to the following unresolved issues:

  1. Circular Validation on Synthetic Dynamics: Even in Section 7, only the schemas and value distributions are imported while all transition rules remain generated by WORLDREL. Evaluating a PFN exclusively on dynamics sampled from its own procedural training prior does not establish transfer to independently authored business logic (e.g., external workflow benchmarks such as World of Workflows, or real SQL/application rules).
  2. Reliability, Consistency, and Calibration Gap: In the stated external API-agent setting—where unsafe trial actions cannot be executed—TabWM still achieves ~0.66–0.67 Edit F1 at K = 64 and incurs ~0.21 normalized regret for H ≥ 2 (ρ ≈ 0.41–0.49). The paper neither guarantees that recursively rolled-out states satisfy database integrity constraints (e.g., foreign-key consistency) nor provides uncertainty calibration/abstention when predictions are unreliable.
  3. Limited Insight for the Community: Beyond engineering an integration of a PFN-style prior, Relational Transformer, and DETR-style set decoder for sparse database edits, the takeaway insight offered to the broader world-modeling and relational-learning community remains below the ICLR bar.

Thank you for recognizing the paper's "sensible and exploratory conceptual direction" and for acknowledging that our additional experiments addressed the horizon concern and improved the paper's positioning (Presentation and Soundness both increased from 2 to 3). We clarify the remaining points below:

1. Evaluation scope.

WORLDREL enables controlled tests of how a model learns hidden relational dynamics from executions. As the original review noted, the lack of aligned real operational traces makes a procedural setting reasonable. Beyond this practical need, controlled worlds let us isolate the evidence that matters.

In §3, we hold the state and action fixed while changing the hidden rules. The resulting edits differ in 65% of paired-world probes. Table 2 tests reassigned action mechanisms, while Appendix F.4 introduces four mechanisms absent from pretraining. These experiments show both adaptation and its limits, particularly for exact value transfer.

Section 7 separately tests real-derived structures and value distributions, not real business rules. We already discuss this limitation in §10 and will clarify the evaluation scope. Beyond addressing the lack of real execution traces, WORLDREL provides a controlled setting for studying how relational models identify and generalize system dynamics. We hope it can also serve as a starting point for further research on this problem.

2. Methodological contribution and insights.

TabWM contributes an in-context learning framework for relational dynamics, not simply another prediction head. Unlike static relational prediction, this task requires learning which changes follow an action in a particular system. WORLDREL varies the dynamics across tasks, the encoder preserves the correspondence between actions and their observed consequences, and the decoder predicts structured edits that can be applied recursively. The numeric transfer ablation further confirms the value of this design (Appendix F.5).

Our experiments reveal several findings beyond the architectural design.

  • Behavioral correspondence matters. Shuffling actions reduces Edit F1 from 0.664 to 0.041 (§6), showing that the model relies on which actions produced which consequences.
  • The model learns more than past edit patterns. With four matching executions, TabWM reaches 0.700 Edit F1 versus 0.216 for direct template transfer (Appendix F.3).
  • Better predictions do not always lead to better decisions. As K increases from 16 to 64, ranking correlation improves from 0.590 to 0.622, while normalized regret remains nearly unchanged (0.208 vs. 0.207, §8).

These findings extend in-context system identification [1] to sparse, multi-table consequences. World of Workflows independently demonstrates the importance of hidden cascading effects in enterprise systems. Our contribution is both a relational model that learns such behavior from experience and a controlled framework for understanding when and how this learning succeeds. We will make these findings and their supporting evidence more prominent.

3. Reliability, consistency, and calibration.

Our planning claim is that learned dynamics improve action choices, not that every simulated state is exact or safe to execute. Section 8 tests this directly. At H=3 and K=64, TabWM achieves 0.207 normalized regret, compared with 0.393 for no-change and 0.392 for shuffled context, a reduction of about 47%. Spearman correlation reaches 0.622 versus 0.345 and 0.338 for these controls.

Prior work also shows that transition accuracy and decision utility are different objectives [2, 3]. Known foreign-key constraints can be checked separately before accepting predicted edits, whereas the hidden transition rules are precisely what TabWM learns from executions. Calibration and abstention are important for autonomous deployment, but our experiments evaluate offline consequence prediction and candidate ranking, not guaranteed safe execution. We will make this distinction clearer.

We hope these clarifications make the scope and significance of our contributions clearer, and we would be truly grateful if you would reconsider your overall rating in light of our responses.

[1] Forgione et al. "From System Models to Class Models: An In-Context Learning Paradigm." IEEE Control Systems Letters, 2023.
[2] Grimm et al. "The Value Equivalence Principle for Model-Based Reinforcement Learning." NeurIPS, 2020.
[3] Lambert et al. "Objective Mismatch in Model-based Reinforcement Learning." L4DC, 2020.
Reviewer_6r5uOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TabWM, a model for predicting action-conditioned changes to relational database states using observed executions from previously unseen systems as in-context evidence. The authors introduce WORLDREL, a generator of executable relational worlds, and formulate relational transition prediction as sparse structured edit prediction. The model combines a relational encoder, a set-based edit decoder, and explicit numeric transformation mechanisms. Experiments demonstrate improvements from behavioral context, robustness to several controlled shifts, and improved short-horizon candidate-action ranking.

Critical strengths

The paper addresses an important gap between conventional relational prediction and modeling the dynamics of operational systems. Its formulation of behavioral in-context learning over relational transitions is compelling and potentially significant. The experimental analysis goes beyond standard predictive benchmarks. In particular, the identical-state/different-dynamics construction, action–consequence shuffling controls, action-map permutations, and action-specific evidence experiments provide insights into what the model learns from observed executions.

The sparse edit representation is appropriate for relational state transitions, and the explicit arithmetic transfer mechanism is supported by a useful ablation. The planning experiments further demonstrate that the learned predictions contain information relevant to downstream action selection.

Critical weaknesses

My primary concern is the generalization boundary. The main held-out-world evaluations resample worlds from the same executable mechanism prior used during pretraining. Although this is a legitimate generalization setting, the results on evaluation-only mechanisms are substantially weaker and include apparent negative transfer with additional context. The evidence therefore supports adaptation within a structured family of relational dynamics more strongly than general identification of previously unseen operational mechanisms. Another concern is the lack of competitive, task-trained structured transition baselines. The flat baselines discard much of the relational and execution-level information used by TabWM, while the LLM baselines do not receive equivalent task-specific training. These comparisons establish performance advantages but do not fully isolate the contributions of the proposed architecture and pretraining strategy.

Also, the action-specific evidence analysis suggests that successful prediction depends strongly on observing the queried action. The paper does not sufficiently establish whether the model can transfer inferred mechanisms between different actions, as opposed to learning action-specific transformation rules from repeated examples.

Although the authors provide world-bootstrap confidence intervals for the main context-length sweep, most central quantitative comparisons are presented as point estimates without corresponding uncertainty measures. This is particularly concerning for the baseline comparisons, distribution-shift evaluations, and planning experiments.

The planning evaluation uses only 40 held-out worlds, with four decision states per world. Since observations from the same world are correlated, uncertainty should be estimated at the world level. Paired world-bootstrap confidence intervals for differences in ranking correlation, normalized regret, and realized return would be important to establish the reliability of the claimed improvements. Moreover, normalized regret changes only from 0.208 at K=16 to 0.207 at K=64, making it difficult to interpret this difference without uncertainty estimates.

A separate concern is training reproducibility. The manuscript does not clearly establish performance variability across independently trained TabWM checkpoints. Resampling evaluation contexts and bootstrapping held-out worlds do not substitute for evaluating multiple training seeds.

Finally, the planning evaluation is restricted to three-step sequences sampled from a pool of oracle-legal candidates and evaluated using a known objective. These results support the utility of the model for short-horizon candidate ranking, but provide limited evidence about robust closed-loop or longer-horizon planning. The treatment of hidden execution state and delayed consequences also warrants clarification.

Questions for authors
  1. What is the model actually learning from behavioral context? The results suggest that performance depends heavily on observing the queried action. To what extent is TabWM identifying transferable world-level dynamics rather than learning action-specific transformations? Can the authors provide evidence that inferred mechanisms transfer to actions not observed in the context?

  2. How well does TabWM generalize beyond its pretraining mechanism distribution? The results on evaluation-only mechanisms are substantially weaker than those on held-out worlds sampled from the same mechanism family. Could the authors clarify the fundamental limitations of this generalization and explain why additional context sometimes fails to improve prediction?

  3. How reliable are the reported improvements? While the paper provides bootstrap confidence intervals for some experiments, several key comparisons, particularly the planning results, lack uncertainty estimates. Could the authors report variability across independent training seeds and paired confidence intervals across evaluation worlds? This would help establish whether the improvements are consistent and statistically reliable.

  4. To what extent does improved transition prediction translate into reliable planning? The planning evaluation considers only three-step trajectories. How does TabWM behave as the planning horizon or search budget increases? In particular, can the authors clarify whether accumulated prediction errors or exploitation of model inaccuracies limit the benefits of additional search?

Please also refer to the Critical Weaknesses section for the main questions and concerns I would like the authors to address.

4 comments

Thank you for the detailed review and for recognizing the value of behavioral in-context learning for relational dynamics. We conducted additional tests of behavioral transfer and planning horizons, and computed paired confidence intervals for the original planning results. We address the main concerns below.

Q1 / W3. What does TabWM learn from behavioral context?

TabWM identifies world-specific action behavior that transfers across relational states, rather than simply replaying observed edits. In Figure 4, shuffling actions reduces Edit F1 from 0.664 to 0.041 at K=64. Table 2 further shows that TabWM adapts to reassigned action-to-mechanism mappings.

We also tested transfer to new relational states. At fixed K=16, increasing matching-action executions from 0 to 4 raises Edit F1 from 0.080 to 0.661 for queries involving previously unbound records, and from 0.075 to 0.657 for cross-table consequences. These results show that inferred action behavior can be applied to different records and relational configurations.

Appendix D.2 (Eqs. 15–16) formalizes behavioral identification in terms of transition rules distinguishable from observed executions. Identifying an unobserved action requires evidence that distinguishes its possible rules, while WORLDREL does not assume that different actions share the same rule. We will clarify this distinction.

Q2 / W1. Generalization beyond the pretraining prior.

Our primary generalization setting is adaptation to unseen systems, not merely unseen database states. Following in-context system identification [1], TabWM learns across a family of systems and adapts to new dynamics through observed executions. Appendix E.2 (Table 8) shows that evaluation worlds have new schemas, action mappings, mechanism parameters, and compositions.

We also evaluate four mechanisms outside the pretraining library (Table 12). In an additional fixed-query analysis, increasing context from K=16 to K=64 improves Support F1 for all four mechanisms, with positive world-bootstrap confidence intervals. The gains include +0.252 for relational copy and +0.280 for linked predicate gating. This shows that behavioral context can help identify which records change even when the exact value transformation is unfamiliar.

These results distinguish the transfer of relational effects from the transfer of exact value transformations, a distinction not captured by Edit F1 alone. We will make this generalization boundary clearer.

W2. Structured baselines and method contributions.

Our evidence for behavioral identification does not rely on the flat baseline comparison alone. Figure 4 isolates the role of action–consequence correspondence, while Table 2 tests adaptation to changed action mappings. Appendix F.3 (Table 11) also evaluates a structured action-template transfer control using the same behavioral context. With four matching-action executions, TabWM achieves 0.700 Edit F1 versus 0.216 for template transfer, showing that its predictions go beyond reusing observed edit patterns. The matched numeric-transfer ablation in Appendix F.5 further improves Edit F1 from 0.476 to 0.616 at K=64.

Flat-TabPFN also retains some relational information through action bindings, link indicators, and summary features (Appendix B.2), but Table 1 is not intended as a matched architectural comparison. Prior relational action-rule learning [2] fits system-specific rules from executions, while TabPFN-Rel [3] uses relational feature synthesis for entity-level prediction. TabWM instead learns a reusable relational transition predictor that adapts to new systems through behavioral context, without parameter updates. We will clarify the roles of these comparisons and controlled experiments.

Q3 / W4–W6. Reliability of the results.

Using the original planning results from Tables 4–5, we computed paired 95% confidence intervals over 40 held-out worlds, keeping all four decision states from each world together.

ComparisonDifference95% CI
Ranking ρ, TabWM − shuffled (K=64)+0.284
+0.199,+0.372+0.199, +0.372

| | Normalized regret, TabWM − shuffled (K=64) | −0.184 | [−0.268, −0.107] | | Mean realized return, TabWM − shuffled (K=64) | +16.882 |

+3.670,+31.100+3.670, +31.100

| | Normalized regret, K=64 − K=16 | −0.0003 |

−0.047,+0.049−0.047, +0.049

| | Median realized return, N=64 − N=16 (K=16) | +0.159 |

+0.009,+0.270+0.009, +0.270

|

The first three comparisons use identical candidate pools and show consistent improvements in ranking, regret, and realized return. As the reviewer noted, regret changes little between K=16 and K=64, while increasing the candidate budget still improves realized return in the tested range.

These intervals measure variation across evaluation worlds, rather than independently trained checkpoints. Appendix F.6 separately examines context resampling. We will clarify the distinction between these sources of uncertainty.

Q4 / W7. Planning horizons and search.

Predicted transitions remain informative across multiple planning horizons, and broader search improves candidate selection in the tested range. We additionally evaluated H=1, 2, 3, and 5 on 40 held-out worlds at K=16 and N=8, using shared candidate prefixes. Spearman correlations range from 0.376 to 0.486, with world-bootstrap confidence intervals above zero at every horizon. The original search-budget experiment further shows improved realized return when increasing N from 16 to 64, as quantified above.

Candidate sequences are generated using action signatures and binding constraints, without filtering through oracle future states. All methods evaluate the same candidate pools. TabWM recursively updates the observable relational state (Eq. 8), while the known objective and action costs (Appendix B.4) are used to score predicted trajectories. We will clarify this protocol and the treatment of hidden execution state in §8.

Thanks again for the valuable feedback and we would like to answer any questions if concern remains. We appreciate your dedicated efforts on reviewing the paper which helps improve the quality of the paper a lot. We are grateful if the reviewer finds our explanation satisfactory and consider re-evaluating the paper.

[1] Forgione et al. "From System Models to Class Models: An In-Context Learning Paradigm." IEEE Control Systems Letters 7 (2023): 3513–3518.
[2] Pasula et al. "Learning Symbolic Models of Stochastic Domains." JAIR 2007.
[3] Hayler et al. "Advancing Open and Reproducible Relational Learning: RelArena-α, TabPFN-Rel and RPI." arXiv preprint arXiv:2608.16319 (2026).
AuthorsOpenReview ↗

Seeking more discussion with Reviewer 6r5u

Dear Reviewer 6r5u,

Thank you again for your thoughtful review. As the discussion period ends on November 18, we wanted to briefly summarize the additional results and clarifications in our response.

  1. We added experiments on previously unbound records and cross-table effects, where Edit F1 improves from 0.080 to 0.661 and from 0.075 to 0.657 with more matching-action examples. We also clarified the structured template baseline comparison (0.700 vs. 0.216) and the contributions of our model components.
  2. We further analyzed four mechanisms absent from pretraining. Increasing K from 16 to 64 improves Support F1 on all four, with positive world-bootstrap confidence intervals. We also clarified why transferring relational effects and predicting exact values present different challenges.
  3. We computed paired 95% confidence intervals over 40 worlds for the original planning results. TabWM improves ranking correlation over shuffled context by 0.284 [0.199, 0.372] and reduces normalized regret by 0.184 [0.107, 0.268].
  4. We added results for H=1, 2, 3, and 5, showing positive ranking correlations across all tested horizons. We also confirmed the benefit of increasing the candidate budget and clarified candidate generation and the rollout protocol.

Thank you again for your valuable suggestions. We look forward to further discussions. Please let us know if we have adequately addressed your concerns, and we are always open to further suggestions. Thank you for your time and consideration.

Authors

I would like to thank the authors for their clarifications and additional experiments. Having read all the reviews, I believe these points strengthen the paper, and I will update my score accordingly.

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

The paper discusses about viewing changes in relational databases from a world modeling perspective: given observable past actions and state changes of relational databases, can a model infer the dynamics behind the actions? To answer this question, the paper builds upon PluRel to obtain WorldRel, which contains synthetic worlds with executable actions and transitions. The paper discusses a wide range of experiments to validate the feasibility of world modeling via relational learning via the synthetic worlds, including checking if the proposed model TabWM can infer the implicit dynamics and decide between candidate action sequences by rollout.

Critical strengths

The strengths of this paper are novelty, informativeness, and experiments. Specifically, the paper executes the idea well via synthetic results, provides clear assumptions/motivations, and isolate alternative hypotheses that are necessary to test in experiments. Below are detailed strengths:

  1. To the best of my knowledge, viewing relational modeling from a world modeling perspective is novel and not discussed previously in literature. The scientific question is interesting and valid.
  2. The experiments support the hypotheses well, with carefully designed experiments to isolate other effects. For example, in Section 4’s Figure 6 the paper isolates improvement from demonstrations vs. improvement from behavioral modeling. The action-consequence prediction improves as number of observations increases, which aligns with the original hypothesis.
  3. Even though TabWM is functionally equivalent to TabPFN, its application to state-action-output tuples seems appropriate to me, and (maybe?) hasn’t been tried previously in literature.
  4. WorldRel can be a valuable asset for ML communities interested in world modeling and relational modeling.
Critical weaknesses

I believe the paper’s clarity can be improved, and the “world modeling” part of the paper may be over-simplified in the synthetic datasets. I identified the following weakness.

  1. While the paper generally has good presentation, I have trouble following the experiments’ settings and their motivations in some parts. For example, Section 6’s motivation seems wordy, and Section 8 introduces action costs and asks audience to read Appendix to understand this concept, even when action cost plays an important role in this section.
  2. I believe the world modeling here is too simplified, since there can be many state changes with unclear actions from the business standpoints. In other words, the paper focuses only on cases where actions are explicit; meanwhile, the state changes are often implicit. To label all actions is thus a big challenge.
  3. The paper only models state-action-output via triplets; meanwhile, world modeling often involves a composition of these triplets to constitute the underlying behavior. In other words, I think the experiments are too simple to be called world modeling.
Questions for authors
  1. In Relational Deep Learning, we often need special care for temporal leakage when training models. I wonder if the authors consider this when training TabWM?
  2. The LLMs used in the paper is Qwen3, which is quite small compared to latest LLM advances. Have the authors compared latest LLMs on the task in Table 1?
3 comments

Thank you for the encouraging review and for recognizing the novelty of our research question and that "the experiments support the hypotheses well"! We clarify the world modeling setting and address your questions below.

W2. Explicit actions and implicit consequences.

Knowing which action was executed does not mean knowing what it changed. We study forward dynamics, where the initiating action is observed but its consequences depend on hidden system rules. World of Workflows [1] studies a similar setting, distinguishing forward audit prediction from action prediction and showing how known operations can trigger hidden cascading changes.

In our Figure 1, expediting one order also affects its shipment, inventory, and another order not mentioned by the action. More importantly, our paired-world experiment (§3, Figure 2) shows that identical states and actions produce different edits in 65% of probes across worlds. The state and action alone therefore do not determine the consequence. Observed executions provide the missing evidence about the active system dynamics. We will make this distinction clearer in §2.

W3. Compositional world modeling.

A world model predicts how actions change states, and these predictions can be composed to imagine future states [2,3]. TabWM supports composition both within an action and across successive actions.

Within an action, WORLDREL applies multiple effects in sequence, allowing changes to propagate across related records. It includes 13 effect classes, with 2.17 effects per action on average and 38.1% of edits occurring outside the action-bound table (Appendix E.1). Across actions, TabWM recursively applies predicted edits to construct future relational states (Eq. 8), and §8 evaluates their use in multi-step action selection. In an additional analysis, Edit F1 on cross-table consequences increases from 0.075 to 0.657 as matching-action evidence grows from 0 to 4 executions at fixed K=16.

Q1. Temporal leakage.

The queried consequence is never available to TabWM at prediction time. RelBench [4] prevents temporal leakage in relational forecasting through timestamp cutoffs. Our setting instead learns dynamics from executions sampled from episodes separate from the query episode. The query provides only its pre-action state and action, with the resulting edits withheld (§2.2, Appendix E.2). Training, development, and evaluation also use disjoint world seed ranges (Appendix C.4). We will clarify this distinction between temporal forecasting and behavioral system identification.

W1. Clarity of the experiments.

We will shorten the motivation in §6 to explain the role of action–consequence correspondence more directly, and introduce the state-potential and action-cost objective in §8 rather than leaving its definition to Appendix B.4.

Q2. Comparison with newer LLMs.

We additionally evaluated GPT-5.6 Luna, a more recent general-purpose API model, on the opaque-naming tasks in Table 1. The results below report Edit F1 across different context sizes.

KQwen3-32BGPT-5.6 LunaTabWM (24.7M)
00.0120.0210.043
40.1270.2400.302
160.3200.4730.524
640.4070.6270.687

At only 24.7M parameters, TabWM achieves higher Edit F1 than both general-purpose models across all four context sizes, providing an additional reference for task-specific relational world modeling.

Thank you again for your helpful feedback!

[1] Gupta et al. "World of Workflows: A Benchmark for Bringing World Models to Enterprise Systems." arXiv preprint arXiv:2601.22130 (2026).
[2] Hafner et al. "Learning Latent Dynamics for Planning from Pixels." ICML 2019.
[3] Pasula et al. "Learning Symbolic Models of Stochastic Domains." JAIR 2007.
[4] Robinson et al. "RelBench: A Benchmark for Deep Learning on Relational Databases." NeurIPS 2024.

Thank the authors for the clarification. I'll keep my positive assessment, and I'd recommend the authors focusing on other reviewers' concerns.

Dear Reviewer KfKq,

We are glad that our response clarified the paper. Thank you for the positive comments.

Thanks!

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Final Remarks

Dear AC and Reviewers,

Thank you all for your time and valuable feedback throughout the review process. We appreciate Reviewer KfKq's recognition of the novelty of our research question and that "the experiments support the hypotheses well", and Reviewer 6r5u's assessment that our formulation is "compelling and potentially significant". We are also grateful that Reviewer L9De raised Presentation and Soundness from 2 to 3 following our responses, and stated that they "would not oppose acceptance" if the AC and other reviewers find the paper's exploratory value sufficient.

We would like to briefly restate what we believe this paper contributes.

Relational learning should go beyond predicting properties of database records. In an operational system, we also need to understand what will change when an action is taken. The difficulty is that the same state and action can produce different consequences under different system rules. We introduce WORLDREL to study this problem and TabWM to learn these hidden dynamics from observed executions, without updating model parameters.

Our results show that TabWM learns from observed behavior rather than simply copying past edits. Identical states and actions produce different consequences in 65% of paired-world probes. Shuffling action–consequence correspondence reduces Edit F1 from 0.664 to 0.041, and TabWM substantially outperforms direct template transfer (0.700 vs. 0.216). The predicted edits can also be used to compare future actions, reducing normalized regret by about 47% compared with the no-change baseline.

During the rebuttal, we further addressed the reviewers' concerns with additional experiments on behavioral transfer, unseen mechanisms, stronger LLM baselines, and different planning horizons. We also provided paired confidence intervals for the planning results and clarified our evaluation setting and method contributions.

We believe this work takes a useful step toward relational world modeling. Beyond the model itself, WORLDREL provides a way to study how relational systems can be understood from execution history, a question that existing static prediction benchmarks do not address. We recognize that evaluation on real operational traces is an important next step, and hope our work provides a useful starting point for further research in this direction.

Thank you again for your careful consideration. We hope the contributions and additional evidence presented during the rebuttal will be taken into account in the final decision.

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.