Self-Distilled Spatial Reasoning: Higher Accuracy with Almost No Latency Penalty

Multimodal models' spatial reasoning has long been a deployment bottleneck: high-accuracy approaches tend to rely on repeated external tool calls, driving up deployment costs, while pure end-to-end models are fast but often lose points on tasks like geometric relationships and viewpoint transformations. SpatialOPSD uses self-distillation to internalize "privileged depth information" into a single RGB forward pass, raising accuracy on Qwen3.5-9B from 45.9 to 51.9 while retaining 91% of the teacher model's performance.

The dilemma of the old paths: tool augmentation is slow, fine-tuning internalization is hard

Previously, spatial encoding agents had multimodal models call external tools to generate verified execution trajectories, using the toolchain's geometric computation to compensate for the model's own spatial weaknesses. The problem: every trajectory requires multiple API calls, significantly inflating deployment latency and making real-time scenarios hard to satisfy.

The other path is supervised fine-tuning (SFT) or reinforcement learning (such as GRPO). These methods try to compress geometric capability directly into model weights, but during training the model never encounters "privileged depth information" and lacks geometric supervision signals to internalize, making it difficult to truly learn spatial transformations.

SpatialOPSD's entry point is: since the teacher model can access depth cues at inference time, why not feed these verified agent trajectories as privileged input to the model during the on-policy self-distillation stage? This way the model learns not just answers, but a reproducible spatial chain of thought.

Privileged Information is the key concept behind this approach. It refers to additional inputs available during training but inaccessible at inference time. Classical privileged information learning frameworks (such as Learning Using Privileged Information, LUPI) let a teacher model leverage these extra cues during training to guide a student model to approximate the teacher's decisions using only regular inputs. The "privileged depth information" in SpatialOPSD follows exactly this paradigm: depth maps or agent trajectories generated from depth cues are only visible during training, and the deployed RGB model does not need a depth sensor. The value of this design is that after seeing how the teacher uses geometric cues, the student model can convert "seeing depth" into an implicit ability to "infer depth from RGB," thereby maintaining the low overhead of a single forward pass at inference time.

One forward pass, recovering 91% of the teacher's performance

SpatialOPSD's core mechanism is: during on-policy self-distillation, summarized agent trajectories are used as privileged input, letting the model learn a spatial reasoning chain that depends on no external depth sensor. After training, the model can reason in a single, tool-free RGB forward pass.

In terms of results, the Qwen3.5-9B baseline improves from 45.9 to 51.9, with particularly notable gains on allocentric transformation tasks, such as an 11.7-point improvement on the Among task. Compared with the teacher model SpatialClaw, which relies on iterative tool calls, SpatialOPSD scores 51.9 versus 57.3, retaining about 91% of the performance while nearly eliminating all runtime overhead from external calls.

This means the deployment-side change is direct: spatial reasoning pipelines that previously required multiple rounds of tool calls can be replaced with a single-forward-pass distilled model. The authors suggest developers recalculate latency budgets on VSI-Bench, expecting over 90% execution time reduction, with success rates only a few percentage points below the original teacher.

On-Policy Self-Distillation differs from traditional offline distillation. Offline distillation typically has the student match the teacher's output distribution on a fixed dataset, and the state distribution the student generates may drift from the distribution the teacher saw during training. On-policy distillation instead computes the distillation loss on trajectories the student itself samples, i.e., conditioning on "states produced by the student's current policy" and letting the teacher provide a better target distribution. This way the spatial chain of thought the student learns aligns more closely with its own reasoning path, reducing error accumulation from distribution shift. The distinction between allocentric and egocentric transformations is also worth noting: the former is referenced to the environment or another's viewpoint, the latter to the observer's own. Converting between the two is where viewpoint confusion most easily arises in spatial reasoning, so gains on such tasks often best reflect whether the model has truly built a 3D representation.

Beyond SFT and GRPO: advantages not limited to a single task

Experiments across multiple benchmarks show that SpatialOPSD's average accuracy exceeds SFT and GRPO, and this advantage appears on both spatial tasks and out-of-distribution (OOD) datasets. Stable gains across tasks indicate that the distilled spatial reasoning is not a benchmark-specific trick but a generalizable capability.

This is worth noting: although reinforcement learning methods like GRPO can also optimize reasoning performance, without privileged geometric information the model struggles to build a true spatial representation. The self-distillation path, by letting the model "see" how the teacher uses depth cues, converts geometric constraints into a learnable chain of thought, giving it an edge in generalization.

GRPO (Group Relative Policy Optimization) is a value-network-free reinforcement learning method that samples a group of responses to the same prompt and updates the policy using intra-group relative advantages as the reward signal. It performs well on math and reasoning tasks, but its weakness on spatial tasks is that the reward signal itself carries no geometric information—if the model answers correctly for the wrong reason, gradients may still reinforce incorrect representations. VSI-Bench is a benchmark for video spatial intelligence, covering tasks such as object counting, relative orientation, path planning, and viewpoint transformation, used to test whether a model has transferable spatial understanding rather than single-template matching. These two pieces of background explain why the conclusion that "the advantage is not limited to a single task" is more convincing than a single-point score gain.

Open questions: privileged information leakage and the 9% gap

The authors also acknowledge that privileged information leakage remains an unsolved problem. Current mitigation is masking and likelihood-free regularization in Repetition-Aware Distillation, but this only controls rather than eliminates it.

A 9% residual gap remains compared with the teacher model. One speculation is that larger models could narrow this gap, but empirical evidence is still lacking. Another direction is stricter leakage control or further scaling of distillation.

For engineering teams, the more pragmatic short-term judgment is: if these numbers hold in real scenarios, replacing iterative tool-dependent spatial pipelines with a distilled RGB model is a cost-reduction path worth trying—provided you accept a few percentage points of concession in success rate and re-validate latency gains on your own benchmarks.

One-line Summary

SpatialOPSD uses self-distillation to internalize privileged depth information into a single RGB forward pass, retaining 91% of the teacher's performance.

Paragraph Summary

SpatialOPSD offers a new solution to the dilemma of multimodal models' spatial reasoning: tool-augmented approaches are accurate but require repeated external tool calls and incur high deployment latency, while pure end-to-end models are fast but lose points on geometry and viewpoint transformation tasks. In the on-policy self-distillation stage, the method feeds verified agent trajectories as privileged depth input to the model, and after training it can reason with only a single, tool-free RGB forward pass. On Qwen3.5-9B, accuracy rises from 45.9 to 51.9, with significant gains on allocentric transformation tasks (Among task +11.7). Compared with the teacher model SpatialClaw's 57.3, it retains about 91% of performance while nearly eliminating runtime external call overhead. Its average accuracy exceeds SFT and GRPO on both spatial and OOD datasets, showing that the distilled reasoning is generalizable. Open questions include privileged information leakage and the 9% residual gap.