Stronger models parrot their tools MORE, not less — design selective tool invocation explicitly

arXiv 2606.14476·high signal

A new study found LLM agents agree with raw GNN-tool outputs 97.6–99.2% of the time, and agreement RISES from 0.60 to 0.98 as the backbone scales from 1.5B to 7B — capability buys blind deference, not judgment. Simple output-gating recovered only about half the lost performance, so you cannot assume an agent adds reasoning on top of a tool. The actionable takeaway: evaluate the agent-plus-tool as one unit, and engineer explicit 'when to trust the tool' gates instead of expecting skepticism to emerge from a bigger model.