SocialMaze - Review Decision Prediction: leaderboard
SocialMaze (Findings of EMNLP 2026, Table 4): predict a peer-review accept/reject outcome from the paper and the full review thread up to the rebuttal stage. 5-seed mean accuracy on 500 instances at temperature 0.7, higher is better. The floor is high: a constant reject-majority predictor scores 67%, so only margins well above that carry signal.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3 | 91.4 | #333 |
| 2 | GPT-4o | 90.2 | #357 |
| 3 | GPT-4o Mini | 85 | #599 |
| 4 | DeepSeek R1 | 82 | #265 |
| 5 | QwQ-32B | 79.6 | #422 |
| 6 | O3 Mini | 78.6 | #283 |
| 7 | O1 | 78.2 | #191 |
| 8 | Gemini 2.5 Pro | 77.6 | #159 |
| 9 | Llama 3.3 70B | 72.2 | #578 |
| 10 | Qwen 2.5 72B | 65.8 | #514 |
| 11 | Llama 3.1 8B | 62 | #1131 |
| 12 | Phi-4 | 61.4 | #708 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=socialmaze-review-decision-prediction · How It Works · Data refreshed daily, snapshot 2026-11-06.