perpetual-research/splatbench

splatbench: RL dataset on Hugging Face. SplatBench is a benchmark and RL environment for raw spatial reasoning. A language model gets eight photos of an object and rebuilds the object in 3D by placing every Gaussian itself through tool calls, with no code and no optimizer. We grade the build from…