While ageing is commonly associated with exponential increase in mortality with age, mortality rates paradoxically decelerate late in life resulting in distinct mortality plateaus. Late-life mortality plateaus have been discovered in a broad variety of taxa, including humans, but their origin is hotly debated. One hypothesis argues that deceleration occurs because the individual probability of death stops increasing at very old ages, predicting the evolution of earlier onset of mortality plateaus under increased rate of extrinsic mortality. By contrast, heterogeneity theory suggests that mortality deceleration arises from individual differences in intrinsic lifelong robustness and predicts that variation in robustness between populations will result in differences in mortality deceleration. We used experimental evolution to directly test these predictions by independently manipulating extrinsic mortality rate (high or low) and mortality source (random death or condition-dependent) to create replicate populations of nematodes, Caenorhabditis remanei that differ in the strength of selection in late-life and in the level of lifelong robustness. Late-life mortality deceleration evolved in response to differences in mortality source when mortality rate was held constant, while there was no consistent response to differences in mortality rate. These results provide direct experimental support for the heterogeneity theory of late-life mortality deceleration.