AI

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

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 09/10/2026, 04:18:24 UTC.

Briefings

A new method called GPD distills spatial reasoning from 3D-scanned training data into vision-language models using geometry as privileged information. On a 4B backbone, GPD improved spatial reasoning benchmarks by 57.1 on VSI-Bench and 37.6 on average across multiple datasets, outperforming standard answer-privileged distillation. The approach renders 3D cues as text alongside answers during on-policy self-distillation, applying a privileged KL loss only to incorrect trajectories while keeping the model RGB-only.

Key points

• GPD distills spatial reasoning from 3D-scanned training data into vision-language models using geometry as privileged information.
• On a 4B backbone, GPD improved spatial reasoning benchmarks by 57.1 on VSI-Bench and 37.6 on average across multiple datasets.

Context

Vision-language models struggle with spatial reasoning because RGB inputs lack geometric evidence. Existing remedies either inject 3D data at inference or train with outcome rewards, which are costly or ineffective. This study proposes GPD, making geometric evidence the privilege in on-policy self-distillation.

Limitations

Evidence is limited to a single 4B backbone model; results may not generalize to larger architectures. The study does not report inference latency costs or architectural trade-offs beyond the 4B backbone. Results are based on benchmark scores rather than user performance or real-world deployment metrics.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 14/11/2026, 11:13:12 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeieihqhfq2gfxcv3zio6whxueoxt2molnrxybq2qb3dmvty3srrf5q