Why LLMs Fail at Causal Discovery and How Interventional Agents Escape: leaderboard
A text-based benchmark on causal discovery using Corr2Cause and Extended Corr2Cause, where models infer causal graphs from observational and interventional information. The evaluated method uses a frozen language model as an interventional oracle within an external Bayesian loop. The pinned metric is F1, measuring graph-prediction quality. Source: BenchmarkList’s attributed result table.
Metric: F1 (%). Source: benchmarklist.com. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.1 | 93.2 | #109 |
| 2 | Qwen 3 30B A3B | 92.1 | #496 |
| 3 | Qwen 3.5 122B A10B | 87.1 | #184 |
| 4 | Llama 3.3 70B | 85.7 | #578 |
| 5 | Gemma 3 12B | 77.6 | #674 |
| 6 | LLaMA-7B | 72.6 | #1543 |
| 7 | GPT-4 | 29.1 | #459 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=why-llms-fail-at-causal-discovery-and-how-interventional-agents-escape · How It Works · Data refreshed daily, snapshot 2026-11-05.