Distilling privileged depth cues into an RGB‑only vision‑language model closes the long‑standing gap between accuracy and inference speed. SpatialOPSD shows that a single forward pass can inherit most of the reasoning power of a tool‑driven teacher without the heavy latency penalty.
Before this work, spatial coding agents boosted multimodal models by calling external tools to generate verified execution traces, but each trace required multiple API calls and slowed down deployment. Fine‑tuning (SFT) or reinforcement‑learning approaches such as GRPO could not internalize geometry because they never saw the privileged depth information during training.
SpatialOPSD improves the Qwen3.5‑9B baseline from 45.9 to 51.9, with specific efficacy in allocentric transformations (e.g., Among +11.7) [1]. The gain comes from feeding summarized agent traces as privileged inputs during on‑policy self‑distillation, which teaches the model a spatial chain‑of‑thought that it can later reproduce without any external depth sensor.
In a single tool‑free forward pass, SpatialOPSD retains 91% of the performance of its iterative, tool‑dependent teacher, SpatialClaw (51.9 vs. 57.3) [1]. This means the distilled model eliminates virtually all runtime overhead from external calls while sacrificing only a modest fraction of accuracy.
Experiments across multiple benchmarks demonstrate that self‑distilling SpatialOPSD achieves higher average accuracy than SFT and GRPO on both spatial and OOD datasets, exhibiting superior performance and generalization [1]. The advantage holds across varied tasks, confirming that the distilled spatial reasoning is not a narrow trick but a broadly useful capability.
The authors acknowledge that privileged‑information leakage remains a concern, mitigated only by Repetition‑Aware Distillation’s masking and unlikelihood regularization; the 9% residual gap to the teacher suggests further work is needed on scaling or tighter leakage control. One inference is that larger models may close this gap, but empirical evidence is still missing.
If these numbers hold in practice, developers should replace iterative tool‑dependent spatial pipelines with a distilled RGB model and re‑run latency budgets on VSI‑Bench; the expectation is a >90% reduction in execution time while keeping success rates within a few points of the original teacher.
Top comments (0)