Counterfactual Reasoning on SCM battery 300 random models per query
0.18Mean Error (PN)diagonal-only (indep. SCM)
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| diagonal-only (indep. SCM)reports the class?=no (single point)2026.06 | 0.18 | — | — | 27 | — | |
| predictor (LLM)reports the class?=no (point, often infeasible)2026.06 | 0.23 | 28 | — | — | — | |
| full kernel (oracle coupling)reports the class?=yes2026.06 | 3 | — | — | — | 0 | |
| full kernel (interval)reports the class?=yes (identified set)2026.06 | — | — | 100 | — | 100 |