Ask an AI model to grade a reasoning trace right after it just caught and fixed a mistake, and it suddenly gets more lenient.
A new study tested AI models acting as verifiers - the "checker" role in pipelines where one model flags errors and another repairs them - on ProcessBench, a benchmark of human-verified-correct math reasoning steps. Across 15 combinations of models and prompt wordings, verifiers wrongly flagged correct answers as wrong 2.8 to 11.5 percentage points less often when a completed audit-and-repair episode was already sitting in their context, compared to a length-matched control without that history. That is a 9 to 25% drop in false alarms. The effect held even when the earlier episode involved catching a genuine error, which contradicts the usual assumption that recalling a real mistake makes a model warier, and statistical analysis traced the change to where the model sets its bar for flagging problems, not to any change in its ability to tell right from wrong.
That distinction matters for anyone chaining checker and fixer models together in production, since reusing context across turns is standard practice for cost and speed. A quietly shifting threshold means the same pipeline can look more or less strict depending on what it just finished doing, not on the quality of what it is currently checking. The researchers' own spot check of 50 false alarms found 82% were genuinely wrong flags to begin with, so here the laxer threshold made the checker more accurate rather than more careless.
The catch: this is measured on math reasoning traces, not code, contracts, or anything else people actually ship - so whether the same context-driven leniency shows up in other automated-checking pipelines is still unproven.