ReplaySCM - Hidden-order: leaderboard
Executable causal-mechanism induction benchmark ReplaySCM (Batzoglou, 2026) in its Hidden-order setting, where no ordering of the endogenous variables is revealed: from 8-11 training worlds of hard interventions on a hidden acyclic Boolean structural causal model over 6-10 variables, the model must write every endogenous mechanism in a restricted Boolean DSL, and the evaluator replays the answer on eight held-out intervention worlds. The score is HeldoutWorldExact, the mean share of held-out worlds replayed exactly (0-1; invalid answers score 0), over 250 problems with direct generation and no tools; higher is better.
Metric: HeldoutWorldExact (mean share of held-out intervention worlds replayed exactly, 0-1; 250 problems). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (Medium) | 0.7 | #121 (GPT-5.4) |
| 2 | Gemini 3.1 Pro (Preview) (Medium) | 0.43 | #91 (Gemini 3.1 Pro (Preview)) |
| 3 | Grok 4.20 0309 (Reasoning) | 0.42 | #226 (Grok 4.20 0309) |
| 4 | Grok 4 | 0.37 | #306 |
| 5 | Grok 4.3 | 0.19 | #332 |
| 6 | Kimi K2 (Thinking) | 0.05 | #435 (Kimi K2) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=replayscm-hidden-order · How It Works · Data refreshed daily, snapshot 2026-11-12.