I think this has a very high chance of success, but trades off the reliability of the probe—if the NLA somehow learns how to hide misalignment semantically (and not steganographically), this type of probe becomes useless.
That would only work if it can reliably find a way to semantically hide misalignment in a way that all transcription models will preserve. That seems unlikely to me.
I think this has a very high chance of success, but trades off the reliability of the probe—if the NLA somehow learns how to hide misalignment semantically (and not steganographically), this type of probe becomes useless.
That would only work if it can reliably find a way to semantically hide misalignment in a way that all transcription models will preserve. That seems unlikely to me.