SpatialOPSD distills depth-based spatial reasoning into RGB-only vision-language models

Distilled spatial reasoning without the tool calls

Researchers behind a method called SpatialOPSD report that a vision-language model that sees only ordinary RGB images can absorb most of the spatial reasoning of a slower, tool-dependent teacher. According to a dev.to writeup summarizing the paper "SpatialOPSD: Self-Distilling Spatial Intelligence from Verified Coding Agent Traces," the distilled model answers in a single forward pass, needs no external depth sensor or extra API calls, and retains 91% of the teacher's measured performance — 51.9 against the teacher SpatialClaw's 57.3.

The gap it tries to close

Earlier attempts to give multimodal models better spatial understanding relied on spatial coding agents: external tools that produce verified execution traces the model can consult. As the writeup notes, every trace costs multiple API calls, which makes the approach awkward to deploy wherever latency matters. Straight fine-tuning and reinforcement-learning schemes such as GRPO failed to internalize geometry, the writeup explains, because the model never saw the privileged depth information during training — only its downstream consequences.

Feeding traces in as privileged input

SpatialOPSD's move is to let the student model see what the teacher sees, but only while training. Summarized agent traces are supplied as privileged inputs during on-policy self-distillation, so the student learns a spatial chain of thought grounded in genuine depth-derived reasoning. At inference time those inputs disappear, yet the reasoning pattern persists: the model reproduces the spatial reasoning from RGB pixels alone, with no tools in the loop and virtually no runtime overhead from external calls.

The reported numbers

• The Qwen3.5-9B baseline climbs from 45.9 to 51.9 under SpatialOPSD, per the writeup.
• Gains concentrate in allocentric transformations, where the "Among" task improves by 11.7 points.
• Across multiple benchmarks, the distilled model beats both SFT and GRPO on average accuracy for spatial and out-of-distribution datasets — evidence, the writeup argues, that the capability generalizes rather than functioning as a task-specific trick.

Caveats the authors flag

The writeup is candid about weaknesses. Privileged-information leakage remains an open concern, addressed only partially through Repetition-Aware Distillation, which combines masking with unlikelihood regularization. The 9% residual gap to the teacher suggests more work on scaling or tighter leakage control is needed. The authors float the possibility that larger models could close that gap but concede that empirical evidence is still missing.

Why it matters

The result matters on two levels. Practically, if the numbers hold outside the lab, teams running iterative, tool-dependent spatial pipelines could swap in a distilled RGB-only model and re-run latency budgets on VSI-Bench; the writeup's expectation is a greater-than-90% reduction in execution time while success rates stay within a few points of the original teacher. For robotics, embodied agents, and on-device multimodal systems, trading a few accuracy points for the elimination of nearly all external-call overhead is an attractive exchange.

Methodologically, SpatialOPSD suggests a broader template: when a capability can be produced with privileged information during training but must be delivered without it at inference, on-policy self-distillation may capture what SFT and reinforcement learning miss. Depth is the demonstration case here, and the same pattern could plausibly extend to other expensive components — sensors, retrievers, simulators — that currently sit inside a model's inference loop.