PERSPECTRA - Opinion Counting: leaderboard
Metric: Exact-match accuracy (%) of the predicted number of distinct underlying opinions in a paragraph that concatenates several expanded arguments, some from the same opinion (PERSPECTRA: Kialo debate opinions (100 topics, 762 pro and con opinions) each expanded into five Reddit-style arguments by GPT-4o; 500 sampled evaluation items per task, gold labels derived from the Kialo opinion identifiers and stances); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | QwQ-32B | 36.2 | #410 |
| 2 | GPT-4o Mini | 34 | #588 |
| 3 | Qwen 3 8B | 33.6 | #667 |
| 4 | DeepSeek-R1-Distill-Qwen-7B | 32.6 | #1400 |
| 5 | DeepSeek R1 Distill Qwen 32B | 31.2 | #640 |
| 6 | Qwen 2.5 7B Instruct | 30.8 | #846 |
| 7 | GPT-4o | 29.2 | #333 |
| 8 | Qwen 3 32B | 28.6 | #424 |
| 9 | DeepSeek R1 Distill Llama 8B | 23.6 | #1282 |
| 10 | Falcon3-7B-Instruct | 15.8 | #1284 |
| 11 | Llama 3.1 8B Instruct | 6.4 | #1018 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=perspectra-opinion-counting · How It Works · Data refreshed daily, snapshot 2026-10-11.