Tool-Use Tax in LLM Agents: When Augmented Reasoning Fails

A new study from arXiv (2605.00136) challenges the consensus that tool-augmented reasoning always improves LLM agents. The authors demonstrate that under semantic distractors, tool-augmented reasoning does not necessarily outperform native chain-of-thought (CoT). They propose a Factorized Intervention Framework to isolate the costs of prompt formatting, tool-calling protocol overhead, and actual execution gains. The analysis reveals a critical tradeoff: under semantic noise, tool gains often fail to offset the 'tool-use tax'—performance degradation from the tool-calling protocol itself. To mitigate this, they introduce G-STEP, a lightweight inference-time gate that partially recovers performance, though more substantial improvements remain needed.