Persian Hate Speech Reasoning - Severity multi-hop: leaderboard
Table 4; temperature 0, top-p 1; multi-hop reasoning except separately named direct one-shot ablation; severity label F1 or impact reference agreement. Metric: Macro F1 (0–100). Higher is better. Source: https://arxiv.org/html/2610.20370v1
Metric: Macro F1 (0–100). Source: arxiv.org. Saturation forecast: Around February 2028. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 50.8 | #109 |
| 2 | Grok 4.6 | 50.7 | #26 |
| 3 | O3 | 44.6 | #123 |
| 4 | DeepSeek V3 | 43.9 | #314 |
| 5 | Claude Sonnet 4.6 | 43.5 | #89 |
| 6 | DeepSeek R1 | 32.1 | #246 |
| 7 | Qwen 3 30B A3B 2507 Instruct | 29.9 | #481 |
| 8 | Qwen 3 30B A3B 2507 (Thinking) | 27.9 | #368 (Qwen 3 30B A3B 2507) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #7 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=persian-hate-speech-reasoning-severity-multi-hop · How It Works · Data refreshed daily, snapshot 2026-10-16.