Chinese Competitive Debating - Winner Prediction: leaderboard
Winner prediction in a benchmark built from 148 real Chinese-language debate matches, each adjudicated by three professional judges under a shared rubric (Yang et al., 2026). The model reads the full match transcript and predicts which side won; accuracy in percent against the majority vote of the judges, higher is better.
Metric: Accuracy (%; the model reads a full match transcript and gives the probability that the affirmative side wins; correct when it picks the side the judges' majority chose, an output of exactly 0.5 counts as wrong; 148 competitive Chinese-language debate matches (4v4 and 2v2 formats), each adjudicated by three professional judges under a shared rubric; zero-shot on manually verified transcripts with fixed prompts; always-affirmative baseline 50.7). Source: arxiv.org. Saturation forecast: Around 2029. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (Medium) | 66.2 | #20 (Claude Opus 5) |
| 2 | DeepSeek V4 Pro | 61.5 | #181 |
| 3 | GPT-5.6 Sol (Medium) | 60.1 | #29 (GPT-5.6 Sol) |
| 4 | DeepSeek V4 Flash (Reasoning) | 60.1 | #254 (DeepSeek V4 Flash) |
| 5 | Claude Sonnet 5 (Medium) | 59.5 | #108 (Claude Sonnet 5) |
| 6 | Gemini 3.5 Flash Lite (Medium) | 59.5 | #300 (Gemini 3.5 Flash Lite) |
| 7 | DeepSeek V4 Flash | 58.8 | #254 |
| 8 | GPT-5.6 Luna (Medium) | 56.1 | #117 (GPT-5.6 Luna) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=chinese-competitive-debating-winner-prediction · How It Works · Data refreshed daily, snapshot 2026-11-12.