SpatialOPSD: Self-Distilling Spatial Intelligence from Verified Coding Agent Traces

SpatialOPSD. Self-Distilling Spatial Intelligence from Verified Coding Agent Traces. By Rongxue Li and eight others, from Alibaba Group. Published on the eighth of October.

Spatial coding agents substantially improve spatial reasoning in multimodal large language models. They call external tools and produce verified execution traces, so each step of the reasoning can be checked. That same design is also their weakness. The inference-time overhead is prohibitive, and the model depends on tools outside itself every time it is asked a question. The capability lives in the system around the model, not in the model. The question the authors pose is whether a model can internalize that agentic capability and run entirely without tools.

Their starting point is a simple observation. When a multimodal model is prompted with summarized execution traces from a spatial coding agent, its own internal spatial chain of thought is unlocked. SpatialOPSD builds on that. It is an on-policy self-distillation framework that treats verified agent traces as privileged information, and uses them to internalize spatial reasoning into a standalone model. On-policy means the model learns from its own attempts, with the verified traces available as extra context that the final model will not have at test time.

Privileged information brings a risk, which is leakage during distillation. To mitigate it, the authors introduce Repetition-Aware Distillation. It combines repetition masking with unlikelihood regularization. The traces guide the model during training, and the regularization is there so the student does not simply carry the privileged material into its own outputs.

Across multiple benchmarks, the self-distilled model reaches higher average accuracy than both SFT and GRPO. That holds on spatial datasets and on out-of-distribution ones, which the authors read as better performance and better generalization.

The arc is from borrowed skill to owned skill. The agent with tools does the verified reasoning once, and the model learns from those traces until it no longer needs the tools, keeping the spatial reasoning while dropping the inference-time cost.