TabWM predicts relational edits from observed executions
Original title: Relational Learning as World Modeling
Across 512 held-out worlds, complete-edit F1 rose from 0.016 with no executions to 0.685 with 64.
The paper introduces TabWM, a tabular world model, and WorldRel, a generator of executable relational worlds for pretraining. Given executions from an unseen world, TabWM predicts sparse structured edits caused by a new action without parameter updates. A correct edit must match operation, relational location, and value. Predicted edits can be applied recursively to compare candidate action sequences before execution.
- Complete-edit F1 is 0.016 with no observed executions and 0.685 with 64.
- Evaluation uses 512 held-out relational worlds.
- Predictions require no parameter updates at test time.
- Broader search over predicted futures improved realized return of selected sequences.
From the original article
1 min read at huggingface.co
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.
By Zheyu Zhang, Gjergji Kasneci · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.