SpatialOPSD: Self-Distilling Spatial Intelligence from Verified Coding Agent Traces
SpatialOPSD distills verified spatial coding-agent traces into a tool-free MLLM that beats SFT and GRPO.
SpatialOPSD is an on-policy self-distillation method that internalizes spatial reasoning from verified coding-agent traces into a standalone multimodal LLM. Summarized traces unlock an internal spatial chain-of-thought, and Repetition-Aware Distillation limits privileged-information leakage with repetition masking and unlikelihood regularization. Across benchmarks, it reports higher average accuracy than SFT and GRPO on spatial and out-of-distribution datasets.
- Internalizes spatial coding-agent traces so an MLLM runs tool-free.
- Treats verified execution traces as privileged information for on-policy distillation.
- Repetition-Aware Distillation uses masking and unlikelihood regularization against leakage.
- Reports higher average accuracy than SFT and GRPO on spatial and OOD sets.
From the original article
1 min read at huggingface.co
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.
By Rongxue Li, Meng Yang, Yiru Mao · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.