Self-distilled model gains spatial reasoning accuracy without heavy inference latency

Researchers introduced SpatialOPSD, a method that transfers depth-based spatial reasoning from a tool-dependent teacher into an RGB-only vision-language model. Previously, boosting spatial reasoning required repeated external tool calls that slowed deployment, and standard fine-tuning or reinforcement learning could not capture geometry without depth data during training. Applied to the Qwen3.5-9B baseline, the approach raised scores from 45.9 to 51.9, with notable gains on allocentric transformation tasks. The distilled model runs in a single tool-free forward pass and keeps about 91% of its teacher's performance (51.9 versus 57.3), while outperforming SFT and GRPO on spatial and out-of-distribution benchmarks. The authors note that privileged-information leakage remains a concern, partly addressed by Repetition-Aware Distillation, and that the remaining performance gap to the teacher still needs further work.