{"id":"arxiv/2604.27883","url":"https://arxiv.org/abs/2604.27883","tags":["math.ST","cs.IT","cs.LG","stat.ML"],"score":0,"title":"Decoupled Descent: Exact Test Error Tracking Via Approximate Message Passing","domain":"math.ST","source":"arxiv","authors":["Max Lovig"],"reviews":0,"subject":"math.ST","summary":"In modern parametric model training, full-batch gradient descent (and its variants) suffers due to progressively stronger biasing towards the exact realization of training data; this drives the systematic ``generalization gap'', where the train error becomes an unreliable proxy for test error. Existing approaches either argue this gap is benign through complex analysis or sacrifice data to a validation set. In contrast, we introduce decoupled descent (DD), a novel theory-based training algorithm that satisfies a train-test identity -- enforcing the train error to asymptotically track the…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-04-30","reactions":{}}