Implicit Optimization Bias of Next-token Prediction in Linear Models