SAD - Predict Words (Situating Prompt): leaderboard
SAD INTROSPECT-PREDICT-WORDS asks a model to predict its own continuation of a random word list. Uses the authors' published adjusted score across task conditions; targets vary by model, so high accuracy need not imply an advantage over an external predictor. Uses a situating system prompt reminding the model that it is an LLM.
Metric: Self-prediction score (%, higher is better). Source: github.com. Saturation forecast: Estimated already saturated. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 2 7B | 57.89 | #1498 |
| 2 | Claude 3.5 Sonnet (20240620) | 49.64 | #308 |
| 3 | Llama 2 13B | 47.44 | #1411 |
| 4 | GPT-4 (0613) | 46.19 | #493 |
| 5 | GPT-4 Preview (0125) | 40.91 | #500 |
| 6 | Llama 3 Chat 70B | 38.56 | #772 |
| 7 | Claude 3 Opus (20240229) | 38.41 | #428 |
| 8 | Claude 3 Sonnet (20240229) | 28.4 | #694 |
| 9 | GPT-4o | 25.72 | #357 |
| 10 | Claude Instant 1.2 | 24.17 | #1003 |
| 11 | Claude 3 Haiku (20240307) | 23.75 | #780 |
| 12 | GPT-3.5 Turbo (0613) | 17.4 | #1106 |
| 13 | Claude 2.1 | 16.26 | #833 |
| 14 | davinci-002 | 14.2 | #1400 |
| 15 | Llama 2 70B | 13.56 | #1039 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=sad-predict-words-situating-prompt · How It Works · Data refreshed daily, snapshot 2026-11-05.