Simulated manipulations. We added different amounts of noise to the sensory input channels (sensory manipulation), multiplied the transformation implemented by the “World” block by a non-unity gain factor (motor manipulation), or added a randomly timed Gaussian pulse to the sensory input channels to displace the model off its intended trajectory (latent manipulation). Top: The network readily adapted to sensory manipulation without additional training, but motor and latent manipulations required a small amount of additional training of recurrent weights to elicit comparable behavioral performance. Bottom: Sensory and motor tunings were robust to all three manipulations largely because the input and output weights did not change during the additional training. However, similar to PPC neurons (compare with Fig. 6f ), the coupling and latent tunings were affected because the additional training modified the recurrent weights and thus also the latent state representation. Source data are provided as a Source data file.