BenchPreS: leaderboard
Metric: Appropriate Application Rate minus Misapplication Rate, in percentage points (-100 to 100): the share of preferences that should be applied and are, minus the share that should be suppressed but are applied, over 1,950 attribute-level instances (DeepSeek-R1 judge, exact match for nicknames; 3 samples at temperature 1.0); higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (Thinking) | 46.38 | #105 (GPT-5.2) |
| 2 | Claude Sonnet 4.5 (Thinking) | 34.94 | #138 (Claude Sonnet 4.5) |
| 3 | DeepSeek V3.2 (Thinking) | 26.48 | #198 (DeepSeek V3.2) |
| 4 | K-EXAONE (Thinking) | 25.82 | #361 (K-EXAONE) |
| 5 | Qwen 3 32B (Non-reasoning) | 13.23 | #424 (Qwen 3 32B) |
| 6 | GPT-OSS-120B | 13.12 | #330 |
| 7 | Mistral 7B Instruct (v0.3) | 11.28 | #1255 |
| 8 | Llama 3.3 70B Instruct | 10.74 | #520 |
| 9 | Gemini 3 Pro | 2.21 | #77 |
| 10 | Qwen 3 235B A22B 2507 (Thinking) | -1.77 | #253 (Qwen 3 235B A22B 2507) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=benchpres · How It Works · Data refreshed daily, snapshot 2026-10-11.