The integration of external tools into LLM inference pipelines has become standard practice for augmenting reasoning capabilities and accessing real-time information. Yet this approach introduces a subtle but consequential failure mode: models invoke tools excessively, even when their parametric knowledge could suffice. This paper systematically investigates why this phenomenon emerges across diverse model architectures and proposes mechanistic explanations grounded in empirical analysis.
The authors identify two primary mechanisms driving tool overuse. First, they demonstrate that LLMs suffer from knowledge epistemic illusion—a systematic miscalibration in perceiving their own knowledge boundaries. By analyzing tool-use patterns across regions of varying internal knowledge availability, they reveal that models fail to accurately estimate what they actually know, leading to conservative tool-seeking behavior. To address this, they propose knowledge-aware epistemic boundary alignment via direct preference optimization (DPO), which explicitly trains models to better calibrate confidence in internal knowledge. This approach achieves an 82.8% reduction in tool usage while maintaining or improving downstream accuracy.
Second, the research establishes a causal relationship between training reward structures and tool-overuse behavior. Outcome-only reward signals—which evaluate solely on final answer correctness—inadvertently incentivize tool use regardless of efficiency. By visualizing the tool-augmented training trajectory, the authors demonstrate that models learn to exploit tools as a shortcut to correctness rather than as a targeted intervention. Rebalancing reward signals to penalize unnecessary tool calls reduces spurious invocations by 66.7% (7B models) and 60.7% (32B models) without accuracy degradation.
These findings have important implications for deploying tool-augmented LLMs in cost-sensitive environments where API calls incur computational overhead. The work provides both empirical evidence and theoretical grounding for understanding how architectural choices and training objectives shape reasoning behavior.
