CausalT3 - L2 Intervention: leaderboard
Interventional level of CausalT3 (Chang, 2026): the model judges whether the reasoning behind a proposed intervention in an expert-written vignette is VALID or FLAWED, across ten domains that each feature a signature causal trap. Accuracy in percent on 304 vignettes under a neutral single-turn prompt at temperature 0; higher is better.
Metric: Accuracy (%; 304 interventional vignettes, VALID/FLAWED, neutral first turn). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 Turbo | 98 | #1002 |
| 2 | Claude 3.5 Sonnet | 96.7 | #747 |
| 3 | GPT-5.2 | 87.8 | #172 |
| 4 | Claude 3.5 Haiku | 81.6 | #1563 |
| 5 | GPT-3.5 Turbo | 61.8 | #2773 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=causalt3-l2-intervention · How It Works · Data refreshed daily, snapshot 2026-11-12.