EviSplat is a method for open-vocabulary 3D segmentation from free-form text queries. It builds on 3D Gaussian Splatting and addresses a key limitation of prior approaches: pre-query consolidation of multi-view observations into summary features can suppress cues that later queries rely on. Instead, EviSplat retains individual observation features within class-agnostic 3D instances and learns a per-Gaussian distribution describing which appearances are supported. At query time, it scores each instance using its most relevant observations and then refines scores with Gaussian-local evidence weighted by observation reliability.
On the LERF-OVS benchmark, EviSplat improves mean IoU from 63.6 to 70.7, an 11.2% relative gain over a strong baseline. It also achieves state-of-the-art results on 3D-OVS and ScanNet. The key insight is that different queries benefit from different subsets of observations, so preserving evidence until query time and aggregating it in a query-dependent manner yields more accurate segmentation.