CodeLift: Your Test Suite Is Secretly a Dense SWE Curriculum
The standard binary reward for training software agents is too sparse, failing to distinguish between partially correct patches. The test pass-rate is better but also unreliable due to imbalances like multiplicity and progress-regression. CodeLift is a reinforcement learning method that dynamically calibrates these rewards by grouping tests and patches based on their behavior, turning the test suite into an adaptive curriculum that provides dense, stable learning signals.