NeSy-Spatial: Self-Evolving Neuro-Symbolic Skills for Spatial Reasoning
A recent submission on arXiv (2608.07955) presents NeSy-Spatial, a neuro-symbolic framework aimed at improving spatial reasoning capabilities in large vision-language models. The researchers highlight that existing models face challenges with detailed spatial tasks that necessitate accurate perception and geometric calculations, which cannot be effectively addressed through end-to-end generation. They suggest tool augmentation as a potential remedy, yet current approaches either plan tool usage without clear dependency constraints or depend on inflexible pipelines that lack generalizability. NeSy-Spatial resolves these challenges by transforming tool interactions and geometric functions into typed executable atomic instructions, categorized into two skill types: Tool-Use Skills for managing tool execution and Geometry Skills for spatial reasoning. This framework supports the development of self-evolving spatial skills, enabling agents to gather reusable experiences and adaptively apply them to new challenges.
Key facts
- Paper arXiv:2608.07955 introduces NeSy-Spatial, a neuro-symbolic framework for spatial reasoning.
- NeSy-Spatial targets large vision-language models' limitations in fine-grained spatial tasks.
- The framework abstracts tool interactions and geometric operations into typed executable atomic instructions.
- It composes two skill types: Tool-Use Skills and Geometry Skills.
- Existing methods for tool augmentation either lack explicit dependency constraints or use fixed pipelines.
- NeSy-Spatial enables self-evolving skills, allowing adaptive composition for new problems.
- The paper is a new announcement on arXiv.
- The approach aims to improve reliability in spatial perception and geometric computation.
Entities
Institutions
- arXiv