Current AI self-evolution systems can't handle vague goals well—they need explicit tasks and metrics to improve. When forced to interpret goals and self-evaluate like humans do, models make poor choices about what data to train on and when to stop, causing improvements to vanish on new tests.
ASPIRE is a benchmark testing whether AI models can self-improve from vague goals (like 'become better at research') without explicit tasks or metrics. Unlike typical self-improvement work, models must interpret goals, choose training data, create their own evaluation signals, and decide when to stop—mimicking how humans learn.