Confidence Decision Faithfulness - GPQA Diamond - Policy Consistency: leaderboard
Whether decisions follow verbal confidence when errors incur penalties. The original paper averages all reported penalty levels within each task. Policy consistency and penalty-normalized regret stay separate; theoretical threshold-policy utility is not an observed model result. Original 0-1 units retained. This series reports Policy Consistency (0-1). Source: https://arxiv.org/html/2601.07767v1.
Metric: Policy Consistency (0-1). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 0.89 | #255 |
| 2 | Gemini 3 Flash | 0.82 | #103 |
| 3 | DeepSeek V3.2 (Non-reasoning) | 0.76 | #210 (DeepSeek V3.2) |
| 4 | GPT-4.1 Mini | 0.76 | #363 |
| 5 | Qwen 3 Next 80B A3B (Thinking) | 0.75 | #312 |
| 6 | Gemma 3n E4B (IT) | 0.72 | #747 |
| 7 | Llama 4 Maverick | 0.71 | #486 |
| 8 | GPT-5 Mini | 0.71 | #193 |
| 9 | DeepSeek V3.2 (Thinking) | 0.69 | #210 (DeepSeek V3.2) |
| 10 | GPT-5 Nano | 0.57 | #439 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=confidence-decision-faithfulness-gpqa-diamond-policy-consistency · How It Works · Data refreshed daily, snapshot 2026-11-05.