EscherVerse: leaderboard
EscherVerse evaluates vision-language models on teleo-spatial reasoning about real-world videos, linking observed spatial change with goal-directed intent. The benchmark is video-based, and the pinned Overall Acc. (%) measures overall answer accuracy on the source leaderboard. Source: BenchmarkList’s attributed result table.
Metric: Overall Acc. (%). Source: benchmarklist.com. Saturation forecast: Around December 2027. 27 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 57.26 | #159 |
| 2 | GPT-5 | 53.36 | #105 |
| 3 | Gemini 2.5 Flash | 52.19 | #255 |
| 4 | Qwen 3 VL 32B (Thinking) | 49.58 | #291 |
| 5 | Qwen 3 VL 30B A3B Instruct | 47.06 | #390 |
| 6 | Qwen 3 VL 4B (Thinking) | 46.58 | #548 |
| 7 | Qwen 3 VL 8B (Thinking) | 46.45 | #423 |
| 8 | Qwen 3 VL 8B Instruct | 46.34 | #419 |
| 9 | Qwen 3 VL 4B Instruct | 45.6 | #510 |
| 10 | InternVL3-8B | 43.56 | #624 |
| 11 | Qwen 2.5 VL 7B Instruct | 38.66 | #650 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=escherverse · How It Works · Data refreshed daily, snapshot 2026-11-05.