Catastrophic failure where a model converges to outputting identical answers regardless of input, despite initial performance gains.