Chinese Social Pragmatic Inference - Full Context: leaderboard
Fine-grained social pragmatic inference over indirect and playful Chinese online comments (sarcasm, humour, wordplay; who is targeted, how the speaker aligns, what the comment does), diagnostic multiple-choice with the full interaction context, leave-writer-out. Higher is better. Source: arxiv.org/abs/2609.04384.
Metric: Leave-writer-out Accuracy (%). Source: arxiv.org. Saturation forecast: Not forecast. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B | 81.42 | #391 |
| 2 | Llama 4 Maverick | 72.59 | #486 |
| 3 | Mistral Medium 3 | 72.18 | #507 |
| 4 | DeepSeek V4 Pro | 71.77 | #112 |
| 5 | MiniMax-M2.5 | 65.89 | #316 |
| 6 | Kimi K2.6 | 63.25 | #113 |
| 7 | GPT-5.5 | 62.95 | #33 |
| 8 | Claude Opus 4.7 | 59.7 | #51 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=chinese-social-pragmatic-inference-full-context · How It Works · Data refreshed daily, snapshot 2026-11-05.