TerraLogic - Spatial Relations Toolkit: leaderboard
Part of TerraLogic (CUHK-Shenzhen and TUM, 2026): 545 scenario-driven geospatial reasoning tasks over optical, SAR and infrared imagery solved through verifiable tool chains, each model driving the HieraPlan reference agent. The score is the end-to-end score of tool-augmented predictions in the spatial-relations toolkit in percent; higher is better.
Metric: End-to-end score (%) of tool-augmented predictions on the Spatial Relations toolkit over the 545 TerraLogic remote-sensing tasks, each model driving the HieraPlan tool agent under a ReAct-style protocol in OpenCompass; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 7B Instruct | 82.42 | #2564 |
| 2 | GPT-4o | 77.95 | #741 |
| 3 | Llama 3 70B Instruct | 62.54 | #1676 |
| 4 | Qwen 2.5 32B Instruct | 61.47 | #1215 |
| 5 | InternLM3-8B-Instruct | 60.7 | #2568 |
| 6 | DeepSeek V3 | 57.93 | #716 |
| 7 | GPT-3.5 | 53.3 | #2417 |
| 8 | Phi-3 Mini 4K Instruct | 51.74 | #6810 |
| 9 | Mistral 7B Instruct (v0.2) | 45.23 | #7007 |
| 10 | Gemini 2.5 Flash | 41.51 | #453 |
| 11 | Yi-1.5-6B-Chat | 36.02 | #6399 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=terralogic-spatial-relations-toolkit · How It Works · Data refreshed daily, snapshot 2026-11-13.