We consider a new variant of successive cancellation decoder (SCD) for polar codes based on the concept of folding, which was proposed in [1], [2] as technique to reduce the decoding latency at the cost of a higher computational complexity. In this paper, we first formally define the multiple folding operation (iterated κ times), which decomposes the original encoding graph into a number of smaller polar encoding graphs. More specifically, we show that the multiple folding gives rise to a two stage interpretation of the graph representing the polar encoder and the SCD. Based on this, we propose the improved multiple folded successive cancellation decoder (IMFSCD), which combines SCD in one stage and maximum-likelihood decoding in the other. This decoder exhibits a latency gain by a factor of 2 κ , still retaining a complexity close to the classic SCD. The small increase in complexity is due to a short maximum likelihood decoder (MLD) used in place of a SCD in the last decoding stage within the IMFSCD. Moreover, we observe by simulation that the decoder performance is exactly the same as that of an SCD at all rates.Keywords-Successive cancellation decoder, multiple folded successive cancellation decoder, partial ML decoding of polar codes, low latency decoder, two stage polar encoder, two stage polar decoder, polar code concatenation.978-1-4799-5999-0/14/$31.00 ©2014 IEEE