Agentic RAG systems enhance LLM through dynamic, multi-stage reasoning and information retrieval, but they can exhibit inefficient retrieval behaviors, such as oversearching (retrieving redundant information) and undersearching (failure to retrieve necessary information). In this study, we define and quantify these behaviors and reveal their prevalence across multiple QA datasets and agentic RAG systems. Furthermore, we identify a significant link between these inefficiencies and uncertainty about the model's knowledge boundaries, revealing that response accuracy is correlated with the model's uncertainty about its retrieval decisions. To address this, we propose $\beta$-GRPO, a reinforcement learning-based training method that incorporates a confidence threshold that rewards high-certainty retrieval decisions. Experiments on seven QA benchmarks show that $\beta$-GRPO enhances the agentic RAG capabilities of the 3B model, outperforming other strong baselines, achieving a 4% higher average accuracy agreement score.
