When producing a token the model doesn't just emit the final token but you also have the entire hidden states from previous attention blocks. These hidden states are mixed into the attention block of future tokens (so even though LLMs are autoregressive where a token attends to previous tokens, in terms of a computational graph this means that the hidden states of previous tokens are passed forward and used to compute hidden states of future tokens).
So no it's not wasteful, those low-perplexity tokens are precisely spots that can instead be used to do plan ahead and do useful computation.
Also I would not be sure that even the output tokens are purely "filler". If you look at raw COT, they often have patterns like "but wait!" that are emitted by the model at crucial pivot points. Who's to say that the "you're absolutely right" doesn't serve some other similar purpose of forcing the model into one direction of adjusting its priors.
Do you know that is true? These aren’t just tokens, they’re tokens with specific position encodings preceded by specific context. The position as a whole is a lot richer than you make it out to be. I think this is probably an unanswered empirical question, unless you’ve read otherwise.
The output is "just tokens"; the "position encodings" and "context" are inputs to the LLM function, not outputs. The information that a token can carry is bounded by the entropy of that token. A highly predictable token (given the context) simply can't communicate anything.
Again: if a tiny language model or even a basic markov model would also predict the same token, it's a safe bet it doesn't encode any useful thinking when the big model spits it out.
train an LLM to leave out the filler words, and see it get the same performance at a lower cost? or do it at token selection time?
Or if you prefer, here's a Galilean thought experiment: gin up a script to get a large language model and a tiny language model to predict the next token in parallel; when they disagree, append the token generated by the large model. Clearly the large model will not care that the "easy" tokens were generated by a different model - how could it even know? Same token, same result. And you will find that the tokens that they agree on are, naturally, the filler words.
To be clear, this observation merely debunks the idea that filler words encode useful information, that they give the LLM "room to think". It doesn't directly imply that an LLM that omits filler words can be just as smart, or that such a thing is trivial to make. It could be that highly predictable words are still important to thought in some way. It could be that they're only important because it's difficult to copy the substance of human thought without also capturing the style. But we can be very sure that what they aren't doing is "storing useful intermediate results".