PCoA - Contributory Phrases: leaderboard
PCoA (Chu et al., University of Duisburg-Essen, 2026) asks a model to summarize one of sixteen medical aspects (such as participants, intervention, outcomes, blinding or funding) of a randomized controlled trial abstract, cite the supporting sentences and extract the contributory phrases from them, over 1,799 expert-annotated instances, zero-shot. This board scores the contributory phrases: token-level F1 (0 to 1) of the phrases extracted from the cited sentences against the reference phrases, counting only phrases that come from a cited sentence and appear in the summary; higher is better.
Metric: Phrase F1 (0-1; token overlap of extracted contributory phrases with the reference phrases, phrases drawn from cited sentences). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Mistral Large 2 (Nov) Instruct (2411) | 0.57 | #1199 |
| 2 | GPT-4o | 0.54 | #741 |
| 3 | DeepSeek V3 (0324) | 0.54 | #704 |
| 4 | Llama 3.1 70B Instruct | 0.54 | #1541 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pcoa-contributory-phrases · How It Works · Data refreshed daily, snapshot 2026-11-12.