Distilling Routed 3D Privilege for Spatial Reasoning in Vision-Language Models

Name: Distilling Routed 3D Privilege for Spatial Reasoning in Vision-Language Models
Uploaded: 2026-10-14T20:18:35.838534+00:00
Duration: 228 s
Description: GPD augments existing spatial reasoning methods by using geometry-privileged self-distillation during training. It combines question-routed 3D cues (depth, semantic, BEV) with the correct answer to guide a student model. This allows the model to learn geometric reasoning while keeping inference RGB-only, outperforming other methods on spatial benchmarks.

GPD augments existing spatial reasoning methods by using geometry-privileged self-distillation during training. It combines question-routed 3D cues (depth, semantic, BEV) with the correct answer to guide a student model. This allows the model to learn geometric reasoning while keeping inference RGB-only, outperforming other methods on spatial benchmarks.

Open the full episode