Epoch AI - Adversarial Nli: leaderboard
Adversarially collected premise-hypothesis pairs crafted to fool strong NLI systems, requiring correct entailment/contradiction/neutrality classification on examples designed to resist pattern matching.
Metric: Score. Source: epoch.ai. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-3.5 Turbo (1106) | 58.1 | #961 |
| 2 | Phi-3-small-8k-instruct | 58.1 | #967 |
| 3 | Llama 3 8B Instruct | 57.3 | #1129 |
| 4 | Phi-3-medium-128k-instruct | 55.8 | #1013 |
| 5 | Mixtral 8x7B (v0.1) | 55.2 | #1083 |
| 6 | Phi-3 Mini 4K Instruct | 52.8 | #1297 |
| 7 | gemma-7B | 48.7 | #1384 |
| 8 | Mistral-7B-v0.1 | 47.1 | #1326 |
| 9 | Phi-2 | 42.5 | #1511 |
| 10 | text-davinci-001 | 35.4 | #1512 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=epoch-ai-adversarial-nli · How It Works · Data refreshed daily, snapshot 2026-10-20.