Skill-3D Builds Scene-Aware Tool-Use Memory for MLLM 3D Spatial Reasoning Agents

Skill-3D addresses uniform tool-use bias in MLLM agents doing 3D spatial reasoning by recording successful tool trajectories into a Scene Memory, then distilling them into reusable scene-specific skills. The framework evolves its skill library based on scene type, improving over both non-agentic baselines and fixed-strategy agents.