This is unironically a technique for catching LLM errors and also for speeding up generation.

For example in speculative decoding or mixture of experts architectures these kind of setups are used.