Self-Evolving Neuro-Symbolic AI Improves Spatial Reasoning with Tools.

Large vision-language models often struggle with precise spatial tasks that demand detailed geometric computation, even when augmented with tools. Existing methods either lack structured planning for tool calls or rely on rigid pipelines that do not generalize effectively. The new NeSy-Spatial framework addresses these limitations by enabling AI agents to learn and adaptively combine reusable skills. NeSy-Spatial operates by abstracting tool interactions and geometric operations into typed executable atomic instructions, forming two complementary skill types: Tool-Use Skills for orchestrating tool execution and Geometry Skills for structured reasoning. During inference, the system retrieves and executes relevant skills in a closed-loop process. Crucially, NeSy-Spatial includes an evolution mechanism that analyzes buffered successful and failed trajectories to refine its skill structures and prune unreliable or inactive entries. This self-improvement process leads to more accurate spatial reasoning and more efficient tool utilization across various benchmarks.