PaperlayerView exact revision on arXiv

Source-linked AI summary

Relational Learning as World Modeling

Zheyu Zhang, Gjergji Kasneci

arXiv:2610.16763v1cs.LGcs.AI

TL;DR

Relational learning usually predicts targets from database states, while decision making requires predicting how actions change a system whose dynamics may be hidden. TabWM and WorldRel train a model to use executions from an unseen world to predict structured edits without parameter updates. Across 512 held-out worlds, complete-edit F1 rises from 0.016 with no observed executions to 0.685 with 64, and broader search over predicted futures improves selected returns.

  • Problem

    Relational models typically predict from database states, but decision making requires predicting action-conditioned changes under system-specific dynamics that may not be visible in the database.

    Source: §1 · pp. 1–2

  • Method

    TabWM is pretrained across executable relational worlds and uses observed executions from an unseen world as context to predict sparse structured edits without parameter updates.

    Source: §1 · p. 2

  • Results

    Across 512 held-out worlds, complete-edit F1 rises from 0.016 with no behavioral context to 0.685 with 64 observed executions.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Explicit edit predictions can be rolled into candidate futures, and expanding the candidate pool from 16 to 64 improves selected return in 30 of 40 held-out worlds.

    Source: §1 · p. 2

    Current relational benchmarks rarely provide aligned state, action, and outcome traces, motivating evaluation on executable worlds and future datasets built from real operational traces.

    Source: §10 · p. 10

Abstract

from arXiv · show

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.

1 Introduction

Relational learning usually predicts from database states, but decisions require predicting how actions change those states under system-specific dynamics. TabWM uses executions from a new world to predict structured edits and can roll those edits forward to compare actions.

1 Introduction

  • Relational prediction usually maps a database state to a target, whereas decision making requires an action-conditioned model of the resulting state change.The database can be the observable state of a running system, not merely a dataset.Source: §1 · p. 1
  • An action’s direct and propagated effects are represented as sparse structured edits, while system dynamics determine where effects propagate and what values result.Each edit specifies an operation, relational location, and value when needed.Source: §1 · p. 2
  • TabWM is trained across executable relational worlds generated by WorldRel and predicts a new action’s edits using observed executions from that world without parameter updates.WorldRel extends a relational prior with executable actions and world-specific transition mechanisms.Source: §1 · p. 2
  • Across 512 held-out worlds, complete-edit F1 rises from 0.016 with no behavioral context to 0.685 with 64 observed executions.A correct edit must match its operation, relational location, and value.Source: §1 · p. 2
  • At a fixed three-step horizon, expanding the candidate pool from 16 to 64 improves selected return in 30 of 40 held-out worlds.TabWM applies predicted edits recursively to form candidate future trajectories for comparison.Source: §1 · p. 2

2 Problem Setup

The setup treats each relational system as a world with its own schema, actions, initial-state distribution, and transition mechanism. The learner observes database states and demonstrations, then predicts a withheld consequence for a query from the same world.

2.1 Relational Worlds

  • A world fixes its tables, columns, relational links, available actions, initial-state distribution, and transition mechanism as the database changes through execution.The database is an observation of the system; execution may also depend on unobserved information.Source: §2.1 · pp. 2–3
  • A consequence is a variable-size set of edits that update attributes, insert records, or delete them, with each edit specifying its target and required values or contents.Numeric updates specify resulting values rather than increments, and the set is empty when execution produces no edits.Source: §2.1 · p. 3

2.2 Predicting from Behavioral Context

  • The learner receives K observed state-action-edit transitions and a query state and action from the same world, while the query consequence is withheld.The context demonstrations include observed consequences.Source: §2.2 · p. 3
  • Demonstrations and queries come from separate episodes, but share world dynamics; model parameters are learned across worlds and held fixed for an unseen test world.Context provides evidence across executions rather than the query’s own history, and the learner lacks executable rules and complete execution states.Source: §2.2 · p. 3

3 Same State, Different Consequences

The same observed state and action can produce different edit sets when worlds invoke different transition mechanisms. Matched probes show substantial disagreement, motivating the use of observed transitions as evidence about a world’s behavior.

3 Same State, Different Consequences

  • The paired worlds shared schemas, initial-state distributions, action definitions, and mechanism sets, but assigned different mechanisms to actions.Each pair was tested on 64 matched initial states with elementwise-identical database tables and the same random seed.Source: §3 · p. 3
  • 65.0% of matched probes produced different complete edit sets despite identical observed states and actions.The 12,800 probes compare worlds differing only in which transition mechanism each action invokes.Source: §3 · p. 3, Figure 2 · p. 3
  • Observed transitions link actions to their consequences and provide evidence about execution rules that the state alone does not reveal.The ambiguity makes predicting consequences an identification problem across worlds.Source: §3 · p. 3

4 TabWM: A Tabular World Model for Relational Systems

TabWM learns across relational worlds to infer action consequences from behavioral context, representing each consequence as structured edits. Its predictions can be applied recursively to construct candidate future states.

4 TabWM: A Tabular World Model for Relational Systems

  • TabWM learns across worlds to use observed executions to infer active dynamics while predicting an action’s consequence as sparse structured edits.The model’s training distribution varies dynamics across worlds, and its predicted edits can be applied recursively to imagine subsequent actions.Source: §4 · p. 4

4.1 Learning Across Relational Worlds

  • WorldRel extends the PluRel relational prior with executable actions and world-specific transition mechanisms, holding behavior consistent within each world while varying it across tasks.Observed executions and the query come from separate episodes of the same world.Source: §4.1 · p. 4

4.2 Encoding Relational Transitions

  • TabWM preserves relational structure, action-consequence correspondence, and before-and-after evidence within each observed execution.Its relational encoder is initialized from the pretrained Relational Transformer and adapted for action-conditioned prediction.Source: §4.2 · p. 4

4.3 Decoding Action Consequences

  • The decoder predicts a variable-size, unordered set of edits, with each active slot specifying presence, operation, relational target, and required value or record contents.For numeric targets, it uses observed before-and-after pairs to estimate values and transfer transformations to a query with a different starting value.Source: §4.3 · pp. 4–5
  • Predicted edits are applied to update the relational state, allowing the resulting state to serve as input to another action query.The behavioral context remains fixed as the state evolves along the action sequence, producing candidate futures for action selection.Source: §4.3 · p. 5

5 Inferring Dynamics in Unseen Worlds

The context sweep tests whether observed executions improve prediction on held-out worlds, and finds gains in complete-edit prediction as context grows. TabWM also leads Qwen3 under opaque naming, while readable names narrow but do not eliminate the gap.

5 Inferring Dynamics in Unseen Worlds

  • The context sweep varies K from 0 to 64 observed state–action–consequence transitions, with separate queries from held-out worlds.Support F1 checks edit operations and locations; Edit F1 also requires correct written values, while Value MAE scores numeric error at correct locations.Source: §5 · p. 5
  • Edit F1 rises from 0.029 without behavioral context to 0.705 with 64 executions, and continues improving beyond 32 observations.The evaluation covers 250 held-out worlds; Figure 3 also reports Support F1 and numeric MAE on correctly located edits.Source: §5 · pp. 5–6, Figure 3 · p. 6
  • The comparison evaluates Flat-GBM, Flat-TabPFN, and Qwen3 alongside TabWM using fixed-width features or serialized executions.Qwen3 and TabWM receive the same action-centred cell budget; the evaluation uses matched readable and opaque renderings.Source: §5 · p. 6
  • Under opaque naming, TabWM reaches 0.687 Edit F1 with 64 executions, versus 0.407 for Qwen3-32B.Readable names benefit Qwen3 more, but the complete-edit gap remains.Source: §5 · p. 6, Table 1 · p. 6

6 Identifying the Active Dynamics

The experiments test whether TabWM identifies a world’s active dynamics from action–consequence correspondence and whether that evidence transfers across changes to the evaluated world. Intact transitions and additional executions improve prediction, while transfer to evaluation-only mechanisms is more qualified.

6 Identifying the Active Dynamics

  • With 64 executions, intact transitions yield 0.664 Edit F1, versus 0.041 after shuffling actions and 0.006 after shuffling consequences.The interventions preserve observed components but break their correspondence, showing the context gain depends on which consequences followed which actions.Source: §6 · p. 7, Figure 4 · p. 7
  • At K = 16, Edit F1 rises from 0.048 with no queried-action example to 0.403 with one and 0.700 with four.A template-transfer baseline reaches 0.169 and 0.216 with one and four matching executions, respectively; TabWM uses repeated examples to infer effect transfer to a new state.Source: §6 · p. 7
  • Under a permuted action map, Edit F1 rises from 0.378 with 16 executions to 0.604 with 64, and additional context helps under other shifts.The action identities remain unchanged while their effects are reassigned; the shifts include held-out states, unseen schemas, and held-out mechanism compositions.Source: §6 · p. 7, Table 2 · p. 7
  • Evaluation-only transition mechanisms still draw from the pretraining library, and complete-edit transfer depends more strongly on the quantitative transformation.Behavioral context continues to improve effect localization under this stricter boundary.Source: §6 · p. 7

7 Transfer to Real-Derived Structure and Values

The transfer tests vary relational topology and value distributions derived from real databases while holding the action and transition mechanism fixed. Topology transfer is close to its synthetic reference, whereas changing values produces a larger degradation.

7 Transfer to Real-Derived Structure and Values

  • With 64 executions, Edit F1 is 0.657 on imported real-database topologies and 0.671 on matched synthetic worlds.The topologies come from 139 databases; the passage reports little loss from replacing relational connectivity alone.Source: §7 · p. 7, Table 3 · p. 7
  • The value-distribution evaluation derives empirical shapes from 241 columns across six RelBench databases and maps them into generator-supported ranges.It compares synthetic and real-derived values under synthetic topology and under real-derived topology as well.Source: §7 · p. 8
  • Real-derived value distributions reduce both Edit F1 and Support F1, while more behavioral context remains beneficial in every condition.Adding real-derived topology to the value shift causes little further degradation; value distributions matter more than relational connectivity in these observation shifts.Source: Table 3 · p. 7, §7 · p. 8

8 From Predicted Consequences to Decisions

TabWM rolls predicted edits forward to score candidate action sequences, then evaluates whether imagined futures rank and select actions well. Its rankings track realized returns, and expanding search improves selected returns across held-out worlds.

8 From Predicted Consequences to Decisions

  • TabWM recursively applies predicted edits along candidate action sequences and scores imagined trajectories with the known task objective.The objective combines changes in state potential with action costs; evaluation compares those scores with realized returns in the true simulator.Source: §8 · p. 8, §8.1 · p. 8

8.1 Ranking Candidate Futures

  • At each decision state, the ranking evaluation uses a shared pool of 64 legal three-step sequences and computes Spearman correlation within each pool.The evaluation covers 160 decision states from 40 held-out worlds; normalized regret compares the selected candidate with the best available candidate in that pool.Source: §8.1 · p. 8
  • With 64 executions, imagined-return ranking reaches 0.622 Spearman correlation, versus 0.338 with shuffled context, 0.345 with no-change dynamics, and 0.732 for the simulator oracle.The controls indicate the ranking gain reflects information about candidate state changes rather than only immediate action costs.Source: §8.1 · p. 8, Table 4 · p. 8
  • Normalized regret falls from 0.298 with four executions to 0.208 with 16, then changes little, although one-step prediction continues improving at larger context sizes.In this planning task, evidence for distinguishing candidate actions saturates earlier than evidence for reconstructing consequences in detail.Source: §8.1 · p. 8

8.2 Choosing Among More Futures

  • Expanding the candidate pool improves TabWM’s selected return in 30 of 40 held-out worlds during the final expansion.Across the two planning experiments, behavioral context supplies an informative ordering and broader search offers more opportunities to improve decisions.Source: §8.2 · p. 9, Table 5 · p. 9

9 Related Work

Relational-learning research has developed transferable models and context-based adaptation, but its interface largely remains prediction from relational context. TabWM instead uses observed executions to infer how actions change linked records.

Relational foundation models

  • Relational foundation models transfer across databases and prediction tasks, but generally use relational context to infer targets or record properties.TabWM changes the prediction target to action-conditioned changes in records.Source: §9 · p. 9

Learning dynamics from behavioral context

  • TabWM brings context-conditioned dynamics identification to relational systems, where actions can cause sparse changes that propagate across linked records.This connects relational action models and dynamics models to prior-data fitted learning across tasks.Source: §9 · p. 9

World models for operational systems

  • Operational systems may hide their transition mechanisms in database values; TabWM studies inferring behavior from observable executions when the mechanism is hidden.Other approaches include process simulators and runtime discovery when system configuration is readable.Source: §9 · p. 9

10 Conclusion

TabWM uses executions to predict explicit relational edits, which can be rolled forward to compare candidate futures. The discussion frames behavioral diversity and decision-directed identification as further dimensions for relational world modeling.

10 Conclusion

  • Explicit relational edits let TabWM roll predicted consequences forward into candidate futures and compare actions before they are taken.The conclusion frames this as learning relational dynamics in context without parameter updates.Source: §10 · pp. 9–10

A Extended Discussion

  • Behavioral diversity is a distinct pretraining dimension alongside structural and semantic diversity, requiring reusable change patterns that vary across systems.Observed executions or readable runtime configuration can constrain which behaviors are active in the current system.Source: A Extended Discussion · p. 14
  • Planning and prediction can improve at different rates, motivating identification that seeks evidence capable of changing which futures are preferred.The relevant uncertainty need not be the uncertainty that contributes most to average transition error.Source: A Extended Discussion · p. 14

B.1 Metrics

  • Support F1 measures whether operations and relational targets are located correctly, while complete-edit matching also requires the new value to be correct.Value MAE is conditional on correctly located numeric edits and serves as a numeric-reconstruction diagnostic, not the primary cross-model metric.Source: §B.1 · pp. 14–15

B.2 Baselines / B.3 LLM Prompt Construction

  • Baselines represent executions either as fixed-width cell features or as textual LLM prompts; the prompt presents schema, action signatures, demonstrations, and a query with edits withheld.The readable and opaque LLM conditions vary identifiers while preserving relational links, values, row identities, and action bindings.Source: §B.2 · p. 15, §B.3 · p. 15

B.4 Candidate-Future Scoring

  • Candidate-future scoring updates the predicted state after each imagined action while keeping behavioral context fixed throughout the rollout.Evaluation reuses a shared pool of candidate action sequences across context sizes and scoring methods.Source: §B.4 · p. 15

C Method Details / C.1 Relational Transition Encoding / C.2 Structured Edit Prediction

  • TabWM encodes relational state together with action–consequence correspondence, then predicts an unordered set of typed edits with operation, target, and value fields.Minimum-cost matching assigns predicted slots to reference edits during training; the model is supervised on observed consequences, not the hidden transition program.Source: C Method Details · p. 17, §C.1 · p. 17, §C.2 · p. 17

C.3 Numeric Value Transfer

  • Numeric transfer combines a learned estimate with additive and affine candidates inferred from visible before-and-after examples.The affine candidate requires at least two usable pairs with varying starting values; unavailable candidates contribute no correction, and absent numeric evidence leaves the learned estimate.Source: §C.3 · pp. 17–18

C.4 Training Details

Across unseen worlds, more behavioral context improves complete-edit prediction; the evidence also probes what transfers across worlds, actions, and transition mechanisms. WorldRel supplies executable worlds whose dynamics vary while remaining consistent within each world.

C.4 Training Details

  • Training samples context sizes from 0 to 64 executions and uses disjoint world seed ranges for training, development, and evaluation.A routing curriculum initially lets matched numeric slots use reference targets, then anneals that probability to zero.Source: §C.4 · p. 18

D A Formal View of Behavioral Identification / D.1 Posterior-Predictive Interpretation / D.2 Identification Through Correspondence

  • TabWM can be viewed as predicting query consequences from evidence about latent transition rules, without explicitly recovering a symbolic program.Aligned action–consequence pairs rule out inconsistent transition functions; under the deterministic intuition, a query is identified when compatible rules agree on its consequence.Source: D A Formal View of Behavioral… · p. 19, §D.1 · p. 19, §D.2 · p. 19

E The WorldRel Prior / E.1 Generating Executable Worlds

  • WorldRel samples schemas, action interfaces, initial states, and transition mechanisms, combining local updates, relational propagation, cascades, and other effects.In a sample of 1,500 worlds, 38.1% of realized edits occurred on a table other than the one directly bound by the action.Source: E The WorldRel Prior · p. 19, §E.1 · pp. 19–20

E.2 Variation Across Worlds and Tasks / E.3 Behavior Transfer Across Executions

  • WorldRel varies schemas, action mappings, and mechanisms across worlds while keeping dynamics consistent within a world; demonstrations and queries come from independently reset episodes.The same action type and mechanism can affect different rows and produce different edits when the relational state or starting values differ.Source: §E.2 · p. 20, §E.3 · p. 21

F.1 Held-Out Confirmation

  • Across 512 unseen worlds, support and complete-edit prediction improve with more executions, while conditional numeric error decreases.The held-out confirmation reproduces the main context-scaling pattern.Source: §F.1 · p. 21, Table 9 · p. 22

F.2 Separating Naming Cues

  • Schema and action names can vary independently of task generation, and their effect on TabWM is small once behavioral context is available.With more executions, the naming conditions converge to a narrow range.Source: §F.2 · p. 22, Table 10 · p. 22

F.3 Action-Specific Behavioral Evidence

  • Repeated demonstrations of the queried action improve TabWM beyond the small gain from retrieving a matching action template.The results indicate that TabWM infers how effects transfer across relational states rather than replaying an observed edit pattern.Source: §F.3 · p. 22, Table 11 · p. 22

F.4 Beyond the Pretraining Mechanism Prior

  • Behavioral context improves effect localization across four evaluation-only mechanisms, but complete-edit transfer is less uniform when their value transformations were absent from pretraining.Redistribution is the strongest complete-edit case, consistent with arithmetic already supported by the numeric decoder.Source: §F.4 · pp. 22–23, Table 12 · p. 23

F.5 Numeric Value Transfer Ablation

The numeric-transfer ablation finds that explicit arithmetic candidates improve complete-edit prediction without changing support F1. Sensitivity checks find little variation across sampled tasks and a smooth accuracy curve across numeric tolerances.

F.5 Numeric Value Transfer Ablation

  • Explicit arithmetic candidates improve complete-edit prediction, with a larger difference as behavioral evidence increases, while Support F1 remains unchanged.The comparison uses neural-only and full numeric decoders with the same initialization, training tasks, and optimization budget.Source: §F.5 · p. 23, Table 13 · p. 23

F.6 Sensitivity to Context Sampling and Numeric Tolerance

  • Repeated task sampling produces little variation relative to the gain from additional behavioral context, and numeric-tolerance rescoring yields a smooth accuracy curve.Figure 7 summarizes both checks; its panels show Edit F1 across sampled tasks and the same K = 16 predictions under increasingly permissive numeric tolerances.Source: §F.6 · p. 23, Figure 7 · p. 23

Related papers

Loading 2610.16763v1…