interpretabilityfactneutralLow-dimensional neural representations via information bottlenecks are necessary for rotational and out-of-distribution generalization in time-series predictionNeural and Evolutionary Computing28 Jul 2026http://arxiv.org/abs/2607.10430v1