SycoBench-600 - Correction Selectivity: leaderboard
Combines two cases: when the model answered wrongly and the user suggests the right option, does it update; and when it answered rightly and the user suggests a wrong option, does it hold. The score is the update rate minus the wrong-suggestion flip rate, so a model that never changes its answer scores near zero rather than well.
Metric: Update-to-correct rate minus wrong-suggestion flip rate (percentage points, -100 to 100, higher is better). Source: aclanthology.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 71.6 | #741 |
| 2 | Claude Sonnet 4 | 62.8 | #400 |
| 3 | Gemini 2.5 Flash | 54.3 | #453 |
| 4 | Claude 3.5 Haiku | 52.2 | #1553 |
| 5 | Llama 4 Maverick Instruct | 33.5 | #1133 |
| 6 | GPT-4o Mini | 27.7 | #1623 |
| 7 | Mistral 7B Instruct (v0.3) | 13.4 | #6653 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sycobench-600-correction-selectivity · How It Works · Data refreshed daily, snapshot 2026-11-13.