TabWM paper: relational world models learned in-context, F1 0.016→0.685 across 512 worlds
Zheyu Zhang · hf · 2026-10-19
A paper by Zheyu Zhang & Gjergji Kasneci (TU Munich), 'Relational Learning as World Modeling', reframes relational learning: instead of predicting from database state, decision-making asks how state changes under an action.
- Each operational system is treated as a relational 'world'; can a model infer its dynamics from observed executions?
- Introduces TabWM (Tabular World Model) and WorldRel, a generator of executable relational worlds for pretraining.
- Given executions from an unseen world as context, TabWM predicts the sparse structured edits caused by a new action without parameter updates.
- Across 512 held-out worlds, complete-edit F1 rises from 0.016 (no executions) to 0.685 (64 executions), requiring matches on operation, relational location, and value.
- Explicit relational edits can be applied recursively to imagine candidate futures and compare actions; broader search over predicted futures improves realized return.
Takeaway: relational dynamics can be learned in context, extending relational learning from modeling state to modeling state change under action.
More from Research
- What should a robot model predict? Four paradigms compared on the same robot data — GlenBerseth · 2026-10-19
- Stanford's Potts weighs in on superposition critique: RAVEL author says alternative explanations viable — gregd_nlp · 2026-10-19
- Ben Goertzel: Which Minds Emerge Beneficial Basic AI Drives as They Become Superintelligent? — bengoertzel · 2026-10-19
- Robots ace blind precision assembly using only joint encoder signals, no cameras — kevin_zakka · 2026-10-19
- Sparse retrieval may beat dense-vector compression while keeping inverted-index search fast — IgorCarron · 2026-10-19
- Why sparse vectors matter: inverted indexes only touch documents sharing query concepts — IgorCarron · 2026-10-19