Risk-Aware LLM Agents for Geospatial Data Retrieval: Design and Preliminary Adversarial Evaluation

Researchers have introduced two new frameworks: one for retrieving remote sensing data using natural language queries and another for optimizing LLM game agents without fine-tuning or human supervision.

The first framework, presented in a paper on arXiv^[1], integrates three agents: Guardrail, General-QA, and Recommender-Analyst, to convert user intent into structured API calls for accessing satellite imagery and environmental datasets. This modular design is portable across platforms and supports applications in environmental monitoring, disaster response, and climate analysis. Preliminary experiments showed that prompt-level safety instructions improve robustness under adversarial conditions. Meanwhile, a second paper on arXiv^[2] introduced an automated prompt optimization framework for LLM game agents. This framework decomposes the observation-to-action pipeline into a goal-conditioned descriptor agent and an action selection agent, iteratively refining each module's prompt through an LLM-driven evolutionary loop. The framework was evaluated on five BabyAI tasks in the BALROG benchmark, achieving a 72.5% success rate on the PutNext task^[2]. In contrast, the RobustCoTAgent achieved a 0% success rate on the same task^[2].