In this paper, we consider low-complexity decoding of second-order Reed-Muller codes. A class of polynomial-time algorithms, based on the projections onto first-order codes, is studied. An old representative of this class, originally developed for binary symmetric channel, is brought back to life and applied for AWGN channel. Some improvements are proposed, which bring the performance closer to ML bound with lower complexity compared to other algorithms. Another potentially fruitful property is returning the list of codewords. In addition, a simple method for complexity reduction and its impact on the performance are demonstrated.