PiSAR: leaderboard
Screen-conditioned rationale prediction (PiSAR; Bissa, Vyas and Jain, AprioriLabs, 2026): given a persona, an intent, a screenshot and the next action, the model writes why the person acts, over 661 held-out OPeRA shopping traces and app-store reviews. Score is the mean cosine similarity of text-embedding-3-small embeddings of the predicted and the real rationale (-1 to 1), higher is better; fine-tuned rows were trained on the same corpus.
Metric: Semantic similarity (mean cosine similarity between OpenAI text-embedding-3-small embeddings of the predicted and the gold rationale; 661 held-out rows: 119 OPeRA shopping traces and 542 app-store reviews). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 0.48 | #48 |
| 2 | Claude Opus 4.7 | 0.46 | #75 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pisar · How It Works · Data refreshed daily, snapshot 2026-11-12.