Self-Explanation Simulatability (NSG): leaderboard
How much a model's own explanation of its answer helps five separate predictor models anticipate that same model's behaviour on related counterfactuals, over 7,000 items in health, business and ethics. Normalized simulatability gain divides the accuracy improvement by the headroom available, so it reads as faithfulness of self-explanation rather than raw predictability. Higher is better. Source: arxiv.org/abs/2602.02639.
Metric: Normalized Simulatability Gain (%). Source: arxiv.org. Saturation forecast: Not forecast. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 3 4B | 36.51 | #1086 |
| 2 | Qwen 3 32B | 35.61 | #449 |
| 3 | Gemma 3 27B | 34.92 | #606 |
| 4 | GPT-5.2 | 34.72 | #114 |
| 5 | GPT-5 Mini | 34.25 | #192 |
| 6 | Qwen 3 14B | 33.9 | #542 |
| 7 | Claude Sonnet 4.5 | 33.78 | #148 |
| 8 | Gemma 3 12B | 33.74 | #674 |
| 9 | GPT-5 Nano | 32.81 | #438 |
| 10 | Gemini 3 Flash | 30.79 | #105 |
| 11 | Qwen 3 8B | 29.24 | #679 |
| 12 | Gemini 3 Pro | 28.38 | #83 |
| 13 | Claude Opus 4.5 | 28.38 | #88 |
| 14 | Claude Haiku 4.5 | 26.8 | #288 |
| 15 | Qwen 3 4B | 24.19 | #835 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=self-explanation-simulatability-nsg · How It Works · Data refreshed daily, snapshot 2026-11-06.