Here's a toy model: self-replicating automatons with no consequence for murder, some automatons murder a percentage of their neighbors' offspring each turn, others do not murder, otherwise they all have the same fertility. Randomly partition the population into two separated groups for a hundred turns, where the two groups have a different incidences of the murdering gene. After the hundred turns, re-combine the two populations, then randomly partition then again.
The murderers will have more surviving offspring within the group, but the group with fewer murderers will grow faster and contribute more genes to the re-combined population. Outcomes are going to depend on specific numbers, and you could make extinction of murderers or extinction of non-murderers more statistically likely.