New tool retrieval method boosts LLM agents for remote sensing tasks

Large language model (LLM)-based agents are increasingly used for automated remote sensing data processing, but they rely on extensive tool libraries whose documentation often exceeds context windows. Existing retrieval methods suffer from 'semantic asymmetry': user queries express high-level goals while tool docs provide fine-grained technical details. A new paper from multiple Chinese universities introduces a bidirectional semantic complementary tool retrieval method to bridge this gap.

On the query side, the method uses a planning-based enhancement mechanism that leverages agent reasoning to break abstract intentions into logical subtasks, actively adding missing functional semantics. On the tool side, it constructs a dynamic tool dependency graph with continual learning, using neighborhood information aggregation to inject precursor tool context into current node representations. Experiments on the remote sensing benchmark GeoPlan-bench and general-purpose API-Bank show significant accuracy improvements and strong extensibility to non-RS domains. The code and dataset are publicly available.