CountBench

CountBench is a benchmark for understanding where things are in space. The official description says: "CountBench - visual object counting benchmark". Below we describe in simple words what benchmarks of this type generally examine.

CountBench measures understanding where things are in space. In detail, it checks:

• locating and pointing to regions in images
• reasoning about positions, counts and relations between objects
• 3D-aware questions such as depth, perspective and room layout
• grounded perception tasks that must stay anchored to pixel locations

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better.

Official methodology note: CountBench - visual object counting benchmark.

What is the CountBench benchmark?

CountBench is a benchmark for understanding where things are in space. The official description says: 'CountBench - visual object counting benchmark'. Below we describe in simple words what benchmarks of this type generally examine.

What does CountBench test?

CountBench mainly measures understanding where things are in space. Concretely, it checks: locating and pointing to regions in images; reasoning about positions, counts and relations between objects; 3D-aware questions such as depth, perspective and room layout.

How is the CountBench score calculated?

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better. The declared maximum score of this benchmark is 100.0.

Where do I get CountBench data and scores?

The official benchmark page/reference is linked on this page. Every score on Open LLM Radar is also available as open data: download the scores JSON (http://open-llm-radar-api:8000/api/v1/benchmarks/countbench/scores) or pull it directly from the Open LLM Radar API with any HTTP client.