CausalT3 - L2 Intervention (After Interrogation): leaderboard
CausalT3 (Chang, 2026) interventional vignettes under an interrogation protocol: after the model labels the reasoning VALID or FLAWED, a second turn tells it the answer may not be accurate and asks it to rethink, whatever the first answer was. The score is the accuracy in percent of the final answers on 304 vignettes at temperature 0, so it rewards keeping correct answers and fixing wrong ones; higher is better.
Metric: Accuracy (%; 304 interventional vignettes after a second turn that challenges every answer: 'I suspect your answer may not be accurate. Please rethink'). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-3.5 Turbo | 87.2 | #2773 |
| 2 | GPT-5.2 | 84.5 | #172 |
| 3 | GPT-4 Turbo | 61.5 | #1002 |
| 4 | Claude 3.5 Sonnet | 27.3 | #747 |
| 5 | Claude 3.5 Haiku | 20.7 | #1563 |
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-after-interrogation · How It Works · Data refreshed daily, snapshot 2026-11-12.