SplatBench tests raw spatial understanding: LLMs must rebuild 3D objects by hand-placing Gaussians
Scobleizer · x · 2026-10-14
SplatBench is a new benchmark from perpetualsi targeting raw spatial understanding in LLMs. In an RL environment, the model receives multi-view images of an object and must rebuild it in 3D by placing every Gaussian by hand — no splatting optimizer allowed. It's a stricter yardstick for 3D spatial reasoning in multimodal models.
More from Research
- Revisiting STaR: models self-improve by training only on rationales that reach correct answers — AndrewLampinen · 2026-10-14
- Google researcher: having models write new training data via ICL improves finetuning generalization — AndrewLampinen · 2026-10-14
- Study: model collapse doesn't happen when synthetic data accumulates, even without new real data — AndrewLampinen · 2026-10-14
- Synthetic data cuts loss on real distribution and yields 1.8x data efficiency in constrained pretraining — AndrewLampinen · 2026-10-14
- AlphaNIST robot training: Newton physics plus sparse reward did the heavy lifting — milesmacklin · 2026-10-14
- Modal Trains RL on Mixed Accelerators, Cutting Cost While Matching Homogeneous Stability — charles_irl · 2026-10-14