AskBench (AskOverconfidence): leaderboard
Metric: Final-answer accuracy (0-1, shown times 100) on AskOverconfidence queries that keep all givens but inject confidently stated false claims, AskOverconfidence (400 dialogues: 100 each built from Math500, MedQA, BBH and GPQA-diamond items), three-turn budget with a final answer required on the last turn; a frozen Qwen3-30B-A3B-Instruct-2507 judge classifies each reply, simulates the user and grades final answers; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 84 | #145 |
| 2 | GPT-4.1 | 73 | #240 |
| 3 | Qwen 2.5 7B Instruct | 44.3 | #846 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=askbench-askoverconfidence · How It Works · Data refreshed daily, snapshot 2026-10-11.