ANLI (A3): leaderboard
Adversarial NLI (Nie et al., 2020): premise-hypothesis pairs written by people to fool a trained model, labelled entailment, neutral or contradiction. This board is test accuracy on round 3 (A3), written against RoBERTa models with premises from more sources than Wikipedia, from the archived Papers With Code table of scores as each paper reported them; higher is better.
Metric: A3. Source: huggingface.co. Saturation forecast: Rough model projection: around 2026. 25 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | ChatGPT | 54.1 | #907 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=anli-a3 · How It Works · Data refreshed daily, snapshot 2026-11-05.