Event-Graph Substrates Enable Exact Counterfactual Reasoning in World Models
A novel category of world models, known as event-graph substrates, employs an append-only log of typed RDF triples to depict agent states and address counterfactual inquiries by forking the log using a structured intervention vocabulary. These substrates allow for triple-level inspection, provide precise counterfactuals, and can be transferred across various domains without requiring learned components. The class has been formalized, demonstrating a duality between explanatory and counterfactual queries, which are both simplified to the same causal-ancestor traversal. A 1,400-line CLEVRER-DSL interpreter was assessed at full validation scale (n=75,618) and outperformed the NS-DR symbolic oracle across all four question categories. Additionally, the paper presents twin-EventLog, a parallel dataset with 500 specifications.
Key facts
- Event-graph substrates represent agent state as an append-only log of typed RDF triples.
- They answer counterfactual queries by forking the log under a structured intervention vocabulary.
- Substrates are inspectable at the triple level and support exact counterfactuals.
- They transfer across domains without learned components.
- A duality between explanatory and counterfactual queries is proved.
- A 1,400-line CLEVRER-DSL interpreter was evaluated on full CLEVRER validation scale (n=75,618).
- The substrate exceeds the NS-DR symbolic oracle on all four per-question categories.
- Improvements over NS-DR: 9.89, 20.26, 17.65, and 0.80 percentage points.
- Exceeds ALOE baseline on descriptive and explanatory but lags on predictive and counterfactual.
- Introduces twin-EventLog, a 500-specification parallel dataset.
Entities
—