ArgBench (Zero-Shot): leaderboard
ArgBench (2026) unifies 46 computational-argumentation tasks from 33 existing datasets into five skills. This board is the macro average over the five skills (mining, perspective assessment, quality assessment, reasoning, and generation scored by BERTScore) under zero-shot prompting, from 0 to 1 times 100; higher is better.
Metric: Macro average (0-1 F1-type scores, times 100) over the five skills (mining, perspective, quality, reasoning, and generation scored by BERTScore) of 46 tasks from 33 existing argumentation datasets, zero-shot prompting; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 | 59.7 | #505 |
| 2 | Llama 3.3 70B Instruct | 56.4 | #1385 |
| 3 | Qwen 3 32B | 53.3 | #972 |
| 4 | Mixtral 8x7B Instruct | 50.8 | #4696 |
| 5 | Llama 3.1 8B Instruct | 49.5 | #4600 |
| 6 | Qwen 3 4B | 47.2 | #2348 |
| 7 | Mistral-Small-Instruct-2409 | 47.1 | #2016 |
| 8 | Phi-mini-MoE-instruct | 46.1 | |
| 9 | Mistral 7B Instruct (v0.3) | 45.3 | #6750 |
| 10 | DeepSeek R1 Distill Qwen 32B | 44 | #1834 |
| 11 | DeepSeek-R1-Distill-Qwen-7B | 40.7 | #8055 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=argbench-zero-shot · How It Works · Data refreshed daily, snapshot 2026-11-21.