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

Hugging Face Daily Papers Papers
spatial-reasoning self-distillation multimodal-llm chain-of-thought coding-agents privileged-information

Summary

SpatialOPSD 提出了一种非策略自蒸馏框架，将空间编码智能体的已验证执行轨迹作为特权信息蒸馏进多模态大模型，通过重复感知蒸馏（重复掩码 + 不似然正则化）缓解信息泄漏，实现无工具的空间推理，并在空间及 OOD 基准上超越 SFT 和 GRPO。

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.
Original Article
View Cached Full Text

Cached at: 10/09/26, 06:25 AM

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

Source: https://huggingface.co/papers/2610.11366

Abstract

SpatialcodingagentssignificantlyimprovespatialreasoninginMultimodalLargeLanguageModels(MLLMs)byusingexternaltoolstogenerateverifiedexecutiontraces.However,thisparadigminherentlysuffersfromprohibitiveinference-timeoverheadandexternaldependencies.Inthispaper,weexplorewhetheranMLLMcaninternalizethisagenticcapabilitytooperateentirelytool-free.Webeginwithasimpleobservation:promptinganMLLMwithsummarizedexecutiontracesofaspatialcodingagentnaturallyunlocksthemodel’sinternalspatialChain-of-Thought(CoT).Motivatedbythis,weintroduceSpatialOPSD,anon-policyself-distillationframeworkthatinternalizesspatialreasoningintoastandaloneMLLMbyformulatingverifiedagenttracesasprivilegedinformation.Tomitigateprivileged-informationleakageduringdistillation,weintroduceRepetition-AwareDistillation,whichcombinesrepetitionmaskingwithunlikelihoodregularization.Experimentsacrossmultiplebenchmarksdemonstratethatself-distillingSpatialOPSDachieveshigheraverageaccuracythanSFTandGRPOonbothspatialandOODdatasets,exhibitingsuperiorperformanceandgeneralization.

View arXiv pageView PDFAdd to collection

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2610.11366 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2610.11366 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2610.11366 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

S-Agent 2.0: Spatial Reasoning with a General Coding Agent

Hugging Face Daily Papers

S-Agent 2.0 通过为通用编码智能体集成命令行感知工具包、持久场景记忆与自然语言解读，显著提升了空间推理能力，在 MMSI-Bench 和 ReVSI 上超越了专用系统 SpatialClaw。

SpatialBrain: Spatial Intelligence Growing from Experience

Hugging Face Daily Papers

SpatialBrain 是一个通过分离场景证据与推理经验来提升空间推理能力的框架，利用融合空间记忆、可复用技能库和过程记录来引导工具使用，且无需修改视觉-语言骨干网络，在五个空间推理基准上取得了优于或持平直接回答的结果。

SAPD: Step-Aligned Privileged Distillation

arXiv cs.CL

论文提出 SAPD（Step-Aligned Privileged Distillation），一种无需 rollout 的自蒸馏方法，将固定示范转化为与推理步骤对齐的分布式监督，在数学推理基准上媲美在线强化学习并实现约 2 倍训练加速。

Pivot-SD: Efficient Self-Distillation for Masked Diffusion Language Models

Hugging Face Daily Papers

Pivot-SD（EMNLP 2026 Oral）提出一种面向掩码扩散语言模型的高效自蒸馏方法，通过信息增益识别关键的"pivot"去噪步骤并仅对这些步骤进行奖励/惩罚，仅用200个问题在数学与代码任务上以更少算力超越全序列SFT和在线RL基线。