ReplaySCM: A Benchmark for Executable Causal Mechanism Induction from Interventions

ReplaySCM introduces a 1,300-item benchmark for evaluating how well language models can infer causal mechanisms from limited intervention data. The benchmark tests whether AI systems can output executable Boolean causal models that generalize to unseen intervention scenarios, revealing that frontier LLMs struggle significantly when structural information is hidden.

ReplaySCM addresses a critical gap in causal reasoning evaluation for large language models. While existing benchmarks focus on scoring local answers or static graph structures, this benchmark demands something more rigorous: the ability to infer and execute functional causal relationships from finite interventional evidence. This distinction matters because correct graph topology doesn't guarantee correct mechanism behavior—a system must output code that actually works on held-out scenarios.

The benchmark's architecture reveals important limitations in current AI capabilities. Frontier models perform reasonably when given ordered structural information but experience sharp degradation in held-out replay performance when order or root structure is hidden. This pattern suggests that LLMs learn superficial causal patterns tied to presentation format rather than developing robust causal understanding. The Alternative-SCM tasks further probe this weakness by asking models to generate semantically distinct mechanisms consistent with training data, then identifying separating interventions.

The counterexample audit ladder—systematically enriching training evidence from original worlds to extra worlds to audited counterexamples—pushes mean predecessor-pattern coverage from 89.49% to 98.15% to 100%. This escalation demonstrates how evidence quality fundamentally constrains what models can reliably infer about causal systems. Even under stronger audit conditions, models still fail to maintain Alternative-SCM consistency with training worlds.

For the AI safety and interpretability communities, ReplaySCM provides actionable methodology for evaluating causal reasoning beyond surface-level metrics. The emphasis on executable replay over formula matching prevents gaming through syntactic variations. This work signals that causal reasoning remains brittle in modern language models and highlights the need for stronger inductive biases or training approaches that build genuine mechanistic understanding rather than pattern matching.

• →ReplaySCM benchmark demands executable causal mechanism inference from finite intervention data, more rigorous than existing causal reasoning evaluations.
• →Frontier LLMs show strong performance in ordered structural settings but degrade sharply when structural information is hidden or roots are unknown.
• →Counterexample auditing raises coverage to 100% but doesn't solve semantic alternative consistency problems in frontier models.
• →Behavior-based scoring prevents syntactic workarounds and ensures mechanisms actually generalize to held-out intervention worlds.
• →Results suggest current language models develop format-dependent causal heuristics rather than robust mechanistic understanding.