Earth-Bench
A benchmark dataset for Earth observation question answering, designed to evaluate the retrieval and reasoning capabilities of agents under complex spatiotemporal constraints. It covers scenarios including vegetation monitoring, disaster assessment and climate change.
Timeline (last 90 days)
Earth-Bench is a benchmark for Earth observation question answering, used to evaluate agents' retrieval and reasoning capabilities under complex spatiotemporal constraints, covering scenarios such as vegetation monitoring, disaster assessment and climate change.
According to the paper, GPT-4o achieves a baseline accuracy of 49.3% on the Earth-Bench benchmark.
On the two benchmarks Earth-Bench and DisasterM3-Reason, OSGR with GPT-4o raised accuracy from 49.3% to 63.7%.
On the two benchmarks Earth-Bench and DisasterM3-Reason, OSGR with GPT-4o reduced the execution step ratio from 2.9 to 2.5.
All Facts (4)
Earth-Bench is a benchmark for Earth observation question answering, used to evaluate agents' retrieval and reasoning capabilities under complex spatiotemporal constraints, covering scenarios such as vegetation monitoring, disaster assessment and climate change.
50%UnverifiedAccording to the paper, GPT-4o achieves a baseline accuracy of 49.3% on the Earth-Bench benchmark.
50%UnverifiedOn the two benchmarks Earth-Bench and DisasterM3-Reason, OSGR with GPT-4o raised accuracy from 49.3% to 63.7%.
50%UnverifiedOn the two benchmarks Earth-Bench and DisasterM3-Reason, OSGR with GPT-4o reduced the execution step ratio from 2.9 to 2.5.
50%