TradeVerse - Statement Generation: leaderboard
Part of TradeVerse (Banerjee and Dasgupta, University of Pisa, 2026), built from World Trade Organization specific trade concern meetings: given every earlier round and the final round without the respondent, the model writes the concluding statement of the responding member. BERTScore F1 in percent against the withheld statement over 1,101 concerns, higher is better; zero-shot at temperature 0.
Metric: BERTScore F1 (%; microsoft/deberta-large-mnli, no baseline rescaling) of the generated concluding statement of the responding member against the withheld final-round statement, given all earlier rounds and the final round's raiser and supporter statements; 1,101 concerns; zero-shot, temperature 0, through the Together API). Source: arxiv.org. Saturation forecast: Around October 2028. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.2 | 60.49 | #138 |
| 2 | DeepSeek V4 Pro | 59.71 | #188 |
| 3 | Kimi K2.7 Code | 57.81 | #185 |
| 4 | Llama 3.3 70B Instruct | 56.67 | #1357 |
| 5 | GPT-OSS-120B | 53.23 | #700 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=tradeverse-statement-generation · How It Works · Data refreshed daily, snapshot 2026-11-13.