ReplaySCM - Block-order: leaderboard
Executable causal-mechanism induction benchmark ReplaySCM (Batzoglou, 2026) in its Block-order setting, where only coarse precedence blocks of the variables are 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 the 100-problem matched pool; higher is better.
Metric: HeldoutWorldExact (mean share of held-out intervention worlds replayed exactly, 0-1; 100 problems). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (Medium) | 0.77 | #121 (GPT-5.4) |
| 2 | Gemini 3.1 Pro (Preview) (Medium) | 0.56 | #91 (Gemini 3.1 Pro (Preview)) |
| 3 | Grok 4.20 0309 (Reasoning) | 0.46 | #226 (Grok 4.20 0309) |
| 4 | Grok 4 | 0.44 | #306 |
| 5 | Grok 4.3 | 0.21 | #332 |
| 6 | Kimi K2 (Thinking) | 0.08 | #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-block-order · How It Works · Data refreshed daily, snapshot 2026-11-12.