CLEM ReferenceGame: leaderboard
CLEM text release v3.0: models identify a target from referring descriptions. Scores combine protocol-valid play and conditional task quality, higher is better. Headline preserves the native published product of macro-average played percentage and macro-average conditional quality; per-game scores multiply source-published played and quality components without extra rounding. Source temperature, player configuration and exact evidenced model/checkpoint settings are retained. Undefined conditional quality remains missing rather than imputed zero.
Metric: Game Clemscore (%), text v3.0; published played × quality / 100. Source: huggingface.co. Saturation forecast: Estimated already saturated. 31 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 100 | #82 |
| 2 | GLM-4.7 | 100 | #192 |
| 3 | GPT-5.2 (High) | 100 | #110 (GPT-5.2) |
| 4 | GPT-OSS-20B | 98.89 | #518 |
| 5 | GPT-5.2 (Medium) | 98.89 | #110 (GPT-5.2) |
| 6 | Kimi K2 (Thinking) | 97.78 | #243 (Kimi K2) |
| 7 | GLM-4.6 | 94.45 | #252 |
| 8 | MiniMax-M2.5 | 86.67 | #303 |
| 9 | Qwen 3 Max | 86.67 | #208 |
| 10 | Mistral Large 3 | 84.44 | #401 |
| 11 | GLM-5 (Non-reasoning) | 76.67 | #141 (GLM-5) |
| 12 | Qwen 2.5 Coder 32B Instruct | 73.33 | #528 |
| 13 | Qwen 3 Next 80B A3B Instruct | 72.22 | #368 |
| 14 | DeepSeek V3.2 (Non-reasoning) | 70 | #205 (DeepSeek V3.2) |
| 15 | Qwen 2.5 72B Instruct | 70 | #454 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=clem-referencegame · How It Works · Data refreshed daily, snapshot 2026-10-20.