The "thinking with images" paradigm often doesn't think about the images.

arXiv 2608.06270 runs a causal audit of visual tool-use, intervening at policy level, trajectory level (corrupting all observations mid-rollout), and step level (counterfactually swapping one observation under a fixed prefix). Across six models and five perception benchmarks: "Calling Without Looking," where returned observations have zero causal effect on the answer, and "Looking Without Planning," where observations are informative but the call schedule is incoherent. Aggregate gains are real but concentrated in a small calibrated minority of rollouts. Most of your crop-and-zoom token spend buys nothing causal.