CausaLab - Graph SHD (5-Node SCMs): leaderboard
The causal graph an LLM agent commits to after its budgeted observations and interventions in CausaLab (Yang et al., 2026), compared with the hidden 5-node structural causal model. Directed structural Hamming distance (missing, extra and reversed edges, a reversed edge counted once), averaged over up to 50 topologies; lower is better.
Metric: Directed structural Hamming distance of the final hypothesised graph to the hidden 5-node SCM (missing, extra and reversed edges; mean per episode, lower is better). Source: arxiv.org. Saturation forecast: Around July 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (High) | 1.32 | #172 (GPT-5.2) |
| 2 | GPT-5 Mini | 3.54 | #349 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=causalab-graph-shd-5-node-scms · How It Works · Data refreshed daily, snapshot 2026-11-12.