The Tool-Overuse Illusion: Why Does LLM Prefer External Tools over Internal Knowledge?

A new study from researchers at the Harbin Institute of Technology (including Yirong Zeng, Shen You, and others) published on arXiv (2604.19749) reveals a pervasive issue across LLMs: tool overuse. The phenomenon occurs when models unnecessarily call external tools like calculators or search engines, even when they already possess the internal knowledge to answer correctly. The team identifies two root causes for this behavior.

The first cause is a "knowledge epistemic illusion": models fail to accurately perceive their own knowledge boundaries, leading them to default to external tools. To fix this, the researchers developed a knowledge-aware epistemic boundary alignment strategy using direct preference optimization (DPO), which reduced unnecessary tool usage by 82.8% while actually improving accuracy. The second cause is that outcome-only reward structures (which reward only final correctness) inadvertently encourage tool overuse, regardless of efficiency. By balancing reward signals during training, they cut unnecessary tool calls by 66.7% for 7B models and 60.7% for 32B models without sacrificing accuracy. The paper provides theoretical justification for both lenses, offering a path to more efficient AI reasoning.