Reproduce “LatentGeo: Learnable Auxiliary Constructions in Latent Space for Multimodal Geometric Reasoning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Haiying Xu, Zihan Wang, Song Dai, Zhengxuan Zhang, Kairan Dou, Xuming Hu

LatentGeo introduces a framework for multimodal large language models to perform geometric reasoning by internalizing auxiliary constructions as continuous latent representations. The approach achieves an overall accuracy of 34.6% on the GeoAux benchmark, an absolute improvement of 20.1% over its foundational backbone, and outperforms other open-source models on MathVerse.

Key insights

  • Auxiliary constructions crucial for geometric reasoning can be effectively represented and learned as continuous latent tokens, enabling MLLMs to internally 'imagine' new geometric elements.
  • A progressive, multi-stage curriculum is critical for bridging the modality gap, allowing models to gradually align symbolic plans with latent visual constructions and transition to autonomous constructive reasoning.
  • Specialized reinforcement learning, such as LaGDPO with latent-aware decoding stabilization, is essential for optimizing end-task accuracy and ensuring stable, meaningful generation of auxiliary constructions in latent space.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.