Does a light rewording of a task hurt an agent as much as a heavy one?

Yes, roughly. Nearly the whole drop happens at the lightest level of rewording, and going heavier from there barely moves the number again. ReliabilityBench tested a no-reword baseline against a light level, synonym swaps, date-format changes, reordered constraints, and a medium level that adds distractor text and mid-task corrections on top. Gemini 2.0 Flash fell from 96.9% success at baseline to 88.1% at the light level, then stayed at 88.1% at medium; GPT-4o fell from 95.0% to 87.5% at light, then edged up to 88.75% at medium rather than falling further.

That’s a different shape than most people expect from “heavier paraphrasing.” An agent’s success rate on a given task isn’t tracking how much the wording changed overall, it’s tracking whether a single sensitive detail got touched at all, and a light reword already touches enough of those for the damage to show up.

It’s a single-author preprint over two models and 1,280 episodes, so treat the shape of the finding as more trustworthy than the exact percentages.