Performance degrades as the number of required function calls increases, making multi-hop reasoning across chained tool calls unreliable. In FuncBenchGen, Megagon Labs frames tool use as traversal over a function-dependency graph and tests seven open and closed LLMs under controlled complexity. GPT-5 drops from 72.5% accuracy with 5 core nodes to 15.0% with 20 core nodes. Connected irrelevant functions that share variables with the solution path further degrade state tracking, causing failures even when individual calls are syntactically valid.
