We propose a new decoding algorithm based on Dror Weitz's tree pruning construction, that interpolates between belief propagation and symbol MAP decoding. We argue that, at low enough noise levels, it provides a polynomial time implementation of symbol MAP decoding.