Формування few-shot прикладів на основі історії дій агента

Abstract

This work proposes an approach to improving LLM-based agents in long horizon planning tasks through dynamic context adaptation and experience driven few-shot learning. Built on the Reason–Act–Observe–Reflect framework with hierarchical task decomposition, the method addresses the inability to provide effective few-shot demonstrations in advance due to dynamic interaction states. The agent accumulates interaction trajectories during execution and converts them into adaptive few-shot examples. This reduces hallucinations and improves performance on complex multi-step tasks.