TRAPSBench: leaderboard
Accuracy on the 202-item Occlusion dataset under the Standard prompt, Table 10. Means over three runs, except a single greedy run for Qwen3-VL 8B. This is not pooled accuracy across all TRAPSBench datasets; thinking and non-thinking checkpoints remain separate.
Metric: Occlusion Accuracy (%, standard prompt). Source: arxiv.org. Saturation forecast: Around January 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 (High) | 82.2 | #98 (GPT-5) |
| 2 | Gemini 3.1 Pro (Preview) (Low) | 82 | #58 (Gemini 3.1 Pro (Preview)) |
| 3 | Gemini 3.1 Flash Lite | 81.4 | #191 |
| 4 | Qwen 3 VL 235B A22B Instruct | 81 | #273 |
| 5 | GPT-5 (Medium) | 80.7 | #98 (GPT-5) |
| 6 | Gemini 3.1 Pro (Preview) | 80 | #58 |
| 7 | Gemini 2.5 Pro (Thinking) | 78.1 | #152 (Gemini 2.5 Pro) |
| 8 | Gemini 2.5 Flash (Non-reasoning) | 76.1 | #245 (Gemini 2.5 Flash) |
| 9 | GPT-5 Mini | 74.8 | #186 |
| 10 | Gemini 2.5 Flash (Thinking) | 73.9 | #245 (Gemini 2.5 Flash) |
| 11 | Qwen 3 VL 235B A22B (Thinking) | 72.3 | #238 (Qwen 3 VL 235B A22B) |
| 12 | Qwen 3 VL 8B Instruct | 58.9 | #420 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=trapsbench · How It Works · Data refreshed daily, snapshot 2026-10-20.