How do distractor functions affect LLM tool use?

Distractor functions are irrelevant functions included alongside the correct solution path during multi-step tool-use evaluation. FuncBenchGen distinguishes two types: connected irrelevant functions that share variables with the solution path but are not needed, and disconnected irrelevant functions that are fully isolated. Connected distractors cause the most damage. They reveal brittle state tracking because the model must distinguish required dependencies from plausible but unnecessary ones. A simple intervention, explicitly restating prior variable values at each step, raised GPT-5 accuracy from 62.5% to 81.3%.