3D Gaussian language maps let robots query a scene with arbitrary text, but when a small set of targets matters in deployment, the distilled features often fail to locate them reliably. This paper proposes a training-free refinement that starts from an already-distilled map and uses a text-promptable segmenter to find each target region. Inside those regions, the visual embedding is blended with the target text embedding via a convex combination; outside, nothing changes.
The key trick is choosing the blending coefficient. Injecting too much prior overwrites appearance cues and hurts general queryability, so the authors solve for the smallest coefficient that still lets the target win the cosine read-out. That coefficient has a closed form, leaving no hyperparameters to tune. They also identify a failure mode: text embeddings share a common component, which can make the read-out query-agnostic after injection. A centering step in the read-out removes this effect.
Across three benchmarks, the refinement consistently improves open-vocabulary segmentation, roughly doubling base mIoU on Replica, and localizes language goals more reliably than the base map. It also transfers as a plug-in to off-the-shelf field-based and object-level language maps, and the authors demonstrate a real-world legged robot navigating to natural-language goals. The work is purely a refinement of existing maps, so it requires no retraining and can be applied to maps already deployed.