SocialMaze - Find the Spy: leaderboard
SocialMaze (Findings of EMNLP 2026, Table 4): each player describes a secret word and the model must identify the one who was given a different word, reasoning over an evolving set of deliberately vague descriptions. 5-seed mean accuracy on 500 instances at temperature 0.7, higher is better. The paper flags this task as confounded with within-family alignment to the LLM pool that generated its synthetic data (subsection I.6), so read cross-family comparisons with care.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 | 78.4 | #191 |
| 2 | Gemini 2.5 Pro | 76.6 | #159 |
| 3 | O3 Mini | 74 | #283 |
| 4 | DeepSeek R1 | 70.2 | #265 |
| 5 | GPT-4o | 69.2 | #357 |
| 6 | GPT-4o Mini | 61.2 | #599 |
| 7 | Llama 3.3 70B | 60 | #578 |
| 8 | DeepSeek V3 | 58.6 | #333 |
| 9 | QwQ-32B | 50.2 | #422 |
| 10 | Qwen 2.5 72B | 48.9 | #514 |
| 11 | Phi-4 | 45.2 | #708 |
| 12 | Llama 3.1 8B | 37.2 | #1131 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=socialmaze-find-the-spy · How It Works · Data refreshed daily, snapshot 2026-11-06.