PCM things are moving and separating
The IEDM paper from IBM  highlighted and detailed some of the very significant "challenges to be overcome" cited by IBM CTO Jai Menon  if the prediction of PCM in IBM servers in 2014 is to be met. The IBM authors  initially acknowledge that both elemental segregation and void formation are the cause of PCM failure.
The fact that material separation occurs during the operation of PCM is not new, it has been known as far back as the 1970s. It has perhaps suffered somewhat from a conspiracy of silence from those promoting other PCM attributes. Bipolar operation has been one of the proposed methods of dealing with it. Alternatively, targeting PCM applications where the possible consequences of shortened write/erase lifetime and/or reduced elevated temperature data retention are acceptable. For most, considering the current densities involved, 10E7A/sq-cm, during PCM reset, the word "electro-migration" was the cover-all to account for the elemental constituents separating and composition change.
More importantly  now provides the evidence of at least two separate mechanisms that can act to change the local composition within the PCM device structure, offering some support to the "never the same device twice" criticism made of PCM devices during write/erase lifetime. Many, this author among them, ,  and  consider that the effect high current density and electro-migration, is the most serious problem standing in the way of the scaling claims for PCM and therefore any future as a mainstream memory product.
From their impressive and admirable piece of TEM/EELS analysis work,  used realistic PCM device structure, the "pore" structure (see Figure 1b). The analysis was focused mainly on devices with 80nm diameter pore. They initially established that with square pulses a much smaller window for "set" was obtained than for pulses with an 80% ramp down. Inconsistent and unacceptable operation was seen for both square pulses and ramp-down pulses when the bottom electrode (BEC) is made positive. Also, it confirmed the known benefits of improved operational characteristics, with respect to the PCM set operation, when a slow ramp rather than a square pulse is employed. That is a slow crystallizing quench. More importantly the observed polarity effect suggested the constrained nature of the asymmetrical structure in the region of the BEC might be the cause.
To test this, the authors  characterized a symmetrical PCM bridge structure, 40nm wide and 200nm long subjected to a single pulse of 3.5v for 5ms.
Initially they confirmed that bridge structures operate best with alternate polarity pulses. It is also perhaps worth noting that these structures operate at high current density . Following a reset to the high resistance state, it was found that a set pulse of the same polarity often leads to a jump to a high resistance state during the ramp down period.
The IBM authors agreed with the results of Yang, Tae Youl et al  that for GST in the molten state, electrostatic force drift is responsible for material separation. Ge and Sb move towards the cathode and Te towards the anode, with effective charge numbers of 0.28, 0.38, -0.24 for Ge, Sb, and Te respectively. The possibility that one atomic species moves by electrostatic force drift or any other means and movement in the opposite direction is by displacement to conserve volume is apparently not considered valid.
Ref  also reported that in crystallized GST at high current density, a more conventional electro-migration effect from a "hole wind force" was responsible for material movement towards the cathode.
Combining the results from pore and bridge structures, the authors in Ref  hypothesized that a structural effect in restricted area devices was the cause of the problem--most likely associated with the build-up tellurium in the confined volume. Therefore negative polarity on the BEC was defined as a good polarity pulse, with a positive BEC as a bad one.