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

Rongxue Li; Meng Yang; Yiru Mao; Yongliang Tao; Lulu Hu; Bin Yang; Zhao Xu; Weihua Luo

cs.LGOct 8, 2026

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

Authors: Rongxue Li, Meng Yang, Yiru Mao, Yongliang Tao, Lulu Hu, Bin Yang, Zhao Xu, Weihua Luo, +1 more

Organizations: Alibaba Group

Abstract

Spatial coding agents significantly improve spatial reasoning in Multimodal Large Language Models (MLLMs) by using external tools to generate verified execution traces. However, this paradigm inherently suffers from prohibitive inference-time overhead and external dependencies. In this paper, we explore whether an MLLM can internalize this agentic capability to operate entirely tool-free. We begin with a simple observation: prompting an MLLM with summarized execution traces of a spatial coding agent naturally unlocks the model's internal spatial Chain-of-Thought (CoT). Motivated by this, we introduce SpatialOPSD, an on-policy self-distillation framework that internalizes spatial reasoning into a standalone MLLM by formulating verified agent traces as privileged information. To mitigate privileged-information leakage during distillation, we introduce Repetition-Aware Distillation, which combines repetition masking with unlikelihood regularization. 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.

Figures & tables

Figure 1 : MindCube-Tiny test accuracy . OPSD from spatial agents effectively boosts spatial reasoning.

Panel of Figure 2 — Figure 2 : (a) Effects of prompting strategies on spatial reasoning. (b) Agent-trace-conditioned CoT increases spatial concept frequency.

Figure 3 : Overview of SpatialOPSD. Verified spatial-agent traces are summarized into privileged information, which guides Repetition-Aware Distillation for spatial reasoning.

Figure 4 : Illustration of Repetition-Aware Distillation.

	Average	MindCube-Tiny				ViewSpatial
	Avg.	Among	Around	Rotation	Overall	CP-	CP-	PP-	PP-	PP-	Overall
						RD	OVO	OVO	RD	SS-RD
Proprietary Model
GPT-5	51.0	38.2	68.4	94.5	56.3	60.2	27.9	41.0	48.5	40.1	45.6
Gemini-3-Pro	60.7	60.7	77.2	93.0	70.9	61.9	31.6	41.1	74.3	38.9	50.4
Open-Source Model

Table 1 : Comparison of the accuracy on spatial reasoning benchmarks. The best results are bolded. CP-RD: Camera perspective relative direction. CP-OVO: Camera perspective object view Orientation. PP-OVO: Person perspective object view orientation. PP-RD: Person perspective relative direction. PP-SS-RD: Person perspective scene simulation relative direction.

Panel of Figure 5 — Figure 5 : Training dynamics of SpatialOPSD-9B and vanilla OPSD, both using Qwen3.5-9B as the backbone.

Table 7

Value	MindCube-Tiny	ViewSpatial	PI Leakage
	Overall (%) $\uparrow$	Overall (%) $\uparrow$	(%) $\downarrow$
Baseline	55.4	46.0	82.7
Varying $\lambda_{\mathrm{rep}}$ ( $L_{\min}=32$ )
0.00	60.0 ( +4.6 )	45.9 ( -0.1 )	73.6 ( -9.1 )
0.05	57.7 ( +2.3 )	47.0 ( +1.0 )	36.0 ( -46.7 )
0.10	56.9 ( +1.5 )	46.8 ( +0.8 )	9.1 ( -73.6 )

Table 4 : Effect of

L_{\min}

and

\lambda_{rep}

on accuracy and privileged information (PI) leakage rate.

Table 9

Panel of Figure 6 — Figure 6 : Training dynamics under different PI variants. We compare FULL-TRACE, SUMMARY, FACT, and INTENT in terms of (a) absolute entropy gap, (b) student entropy, and (c) teacher entropy.

Table 11

Figure 7 : Performance comparison of Qwen3.5-9B-GRPO and SpatialOPSD-9B on MindCube-Tiny.

Figure 8 : Privileged information leakage. The baseline generates unexecutable agent-like traces, whereas SpatialOPSD-9B produces grounded spatial reasoning.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Variant	Constructor	Privileged body	Answer stated	Teacher images
Full-trace	Deterministic	Raw execution log	Yes	Agent show() frames
Summary	Qwen3.5-9B, style=full	Key reasoning points	No	Cited images
Fact	Qwen3.5-9B, style=facts	Tool-measured facts	Appended	Cited images
Intent	Qwen3.5-9B, style=intent	Tool-free strategy	No	None

Appendix

Table 9: The four privileged-information variants, all constructed from the same

5{,}008

agent-correct SpatialClaw traces. “Teacher images” counts what the teacher receives in addition to the original input frames that the student also sees.

Figure 9 : Category-level lexical frequency before and after spatial-reasoning training. Frequencies are normalized by rollout length and reported as occurrences per 1,000 words. The value above each bar gives the exact plotted frequency.

Figure 10 : Rollout coverage before and after spatial-reasoning training. Coverage is the percentage of the 1,040 examples whose rollout contains at least one marker from the corresponding category. The value above each bar gives the exact plotted percentage.

Category	Lexical markers
Step Numbering	step 1 , step 2 , step 3
Multi-Image Comparison	compare , both images , using the three
Initial State Reference	original , initial
Verification and Conclusion	understand , verify , conclusion
Scene Context Analysis	scene , context , situation
Predictive Simulation	if the camera , into view , would bring , would make , will bring , would move , brings you closer , would shift , would appear

Appendix

Table 10 : Lexical markers used to define each reasoning category.

Explore similar work

Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes
Oct 1, 2026Sophia Sirko-Galouchenko, Monika Wysoczanska, Andrei Bursuc +2VLM Distillation Multimodal Large Language Models
Sorbonne Universit´e, CNRS, ISIR, F-75005 Paris, France · ILLS, CNRS, Montreal, QC H2S 3H1 · Institut universitaire de France (IUF)
CoCoSI: Collaborative Cognitive Map Construction for Spatial Intelligence
Jun 9, 2026Yiming Zhang, Ruoxuan Cao, Zhihang ZhongMultimodal Large Language Models Multi-Agent LLMs
2Cornell University · 1Shanghai Jiao Tong University
Spatial-OPSD: Self-Improving Spatial Reasoning via Label-Free Self-Distillation
Sep 29, 2026Zhenyu Liu, Zhangquan Chen, Keyi Chen +4Visual Spatial Reasoning 3D Spatial Reasoning
Tsinghua University · LMMs-Lab · Xi’an Jiaotong University

Oct 1, 2026, cs.CVJ/K move · Enter open · S save

Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes

Sophia Sirko-Galouchenko, Monika Wysoczanska, Andrei Bursuc, Nicolas Thome+1

Sorbonne Universit´e, CNRS, ISIR, F-75005 Paris, France · ILLS, CNRS, Montreal, QC H2S 3H1 · Institut universitaire de France (IUF)

VLM Distillation Multimodal Large Language Models

Open paper AskPDF arXiv

On-policy self-distillation has recently emerged as an effective approach for improving language-model reasoning by supervising students with a frozen or EMA version of themselves that receives privileged information. Its application to multimodal large language models (MLLMs), however, remains largely unexplored. Recent approaches use privileged visual information, such as image crops corresponding to a question, to improve fine-grained perception, but their gains are confined to tasks that benefit from such visual zooming and require either human-annotated grounding data or external teacher models. We introduce a different form of on-policy self-distillation for MLLMs that provides the teacher with textual, spatially grounded guidance identifying the visual elements relevant to a query. We use procedurally generated scenes with automatically available object identities and spatial coordinates, enabling scalable and annotation-free post-training. The teacher uses this spatial guidance to locate and integrate evidence from multiple relevant image regions, while the student learns to reproduce the resulting behavior from the image and question alone. Our approach consistently improves performance on counting, document and chart understanding benchmarks across multiple models. Importantly, although post-training uses only synthetic scenes, the resulting improvements transfer to real-world perception benchmarks, yielding a 3.23-point gain in average performance across CVBench, V*, ZoomBench, BLINK, HR-Bench, and MME-RealWorld. These results show that spatially grounded privileged information can induce broader perceptual capabilities through on-policy self-distillation, enabling substantial synthetic-to-real transfer beyond the task and data distribution used for post-training. Project page: https://github.com/sirkosophia/Where-OPD

Abstract

Explore similar work

Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes

CoCoSI: Collaborative Cognitive Map Construction for Spatial Intelligence

Spatial-OPSD: Self-Improving Spatial Reasoning via Label-Free Self-Distillation