This paper presents a new method for soft decision decoding of binary linear block codes. An efficient algorithm for the numerical computation of soft outputs given by an AWGN channel is developed and investigated based on a special nonlinear . regression model. The minimization of the sum of squared errors in the considered regression model leads to an unconstrained nonlinear optimization problem with continuously differentiable objective function , which is numerically solved by the BFGS-method. Numerical results using BCH codes are summarized.