SpatialOPSD
A spatial reasoning model based on Online Policy Self-Distillation (OPSD). It distills privileged 3D geometric cues into text-based chain-of-thought, enabling the model to perform spatial relationship reasoning using only RGB image input without access to depth or point-cloud data.
Timeline (last 90 days)
A 9% residual gap remains between SpatialOPSD and the teacher model; the authors speculate that a larger model might narrow this gap, but empirical evidence is currently lacking.
SpatialOPSD internalizes privileged depth information into a single RGB forward pass through self-distillation, raising spatial reasoning accuracy from 45.9 to 51.9 on Qwen3.5-9B.
SpatialOPSD retains about 91% of the performance of the teacher model SpatialClaw on Qwen3.5-9B (51.9 versus 57.3).
SpatialOPSD shows especially notable gains on allocentric transformation tasks, for example improving by 11.7 points on the Among task.
The privileged depth information in SpatialOPSD (depth maps or agent trajectories generated from depth cues) is visible only during training, and the RGB model at deployment does not need a depth sensor.
Experiments on multiple benchmarks show that SpatialOPSD's average accuracy is higher than SFT and GRPO, and this advantage appears simultaneously on spatial tasks and out-of-distribution (OOD) datasets.
Privileged information leakage remains an unsolved problem for SpatialOPSD; the current mitigation is masking and likelihood-free regularization in repeated perception distillation, but this only controls rather than eliminates it.
SpatialOPSD uses validated programming agent execution trajectories and geometric descriptors to train a model that relies solely on RGB input, surpassing the SFT and GRPO baselines on the VSI-Bench benchmark.
All Facts (8)
Privileged information leakage remains an unsolved problem for SpatialOPSD; the current mitigation is masking and likelihood-free regularization in repeated perception distillation, but this only controls rather than eliminates it.
50%UnverifiedSpatialOPSD internalizes privileged depth information into a single RGB forward pass through self-distillation, raising spatial reasoning accuracy from 45.9 to 51.9 on Qwen3.5-9B.
50%UnverifiedSpatialOPSD retains about 91% of the performance of the teacher model SpatialClaw on Qwen3.5-9B (51.9 versus 57.3).
50%UnverifiedSpatialOPSD shows especially notable gains on allocentric transformation tasks, for example improving by 11.7 points on the Among task.
50%UnverifiedThe privileged depth information in SpatialOPSD (depth maps or agent trajectories generated from depth cues) is visible only during training, and the RGB model at deployment does not need a depth sensor.
50%UnverifiedExperiments on multiple benchmarks show that SpatialOPSD's average accuracy is higher than SFT and GRPO, and this advantage appears simultaneously on spatial tasks and out-of-distribution (OOD) datasets.
50%UnverifiedSpatialOPSD uses validated programming agent execution trajectories and geometric descriptors to train a model that relies solely on RGB input, surpassing the SFT and GRPO baselines on the VSI-Bench benchmark.
50%UnverifiedA 9% residual gap remains between SpatialOPSD and the teacher model; the authors speculate that a larger model might narrow this gap, but empirical evidence is currently lacking.
50%