SuperGLUE COPA: leaderboard
Choice of Plausible Alternatives (Roemmele et al., 2011), a SuperGLUE task: given a premise, pick which of two alternatives is its more plausible cause or effect. Archived Papers With Code table of accuracies as each paper reported them, under that paper’s own setup; higher is better.
Metric: Accuracy. Source: huggingface.co. Saturation forecast: Rough model projection: around 2026. 60 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | PaLM 540B (finetuned) | 100 | |
| 2 | Vega v2 6B (KD-based prompt transfer) | 99.4 | |
| 3 | ST-MoE-32B 269B (fine-tuned) | 99.2 | |
| 4 | UL2 20B (fine-tuned) | 99 | |
| 5 | DeBERTa-Ensemble | 98.4 | |
| 6 | Turing NLR v5 XXL 5.4B (fine-tuned) | 98.2 | |
| 7 | DeBERTa-1.5B | 96.8 | |
| 8 | PaLM 2-L (1-shot) | 96 | |
| 9 | T5-XXL 11B (fine-tuned) | 94.8 | |
| 10 | FLAN 137B (prompt-tuned) | 94 |
Interactive version: theaggregate.ai/benchmark?slug=superglue-copa · How It Works · Data refreshed daily, snapshot 2026-11-05.