Reminded me of this paper from last year trying to optimize efficient token usage providing budget guidance information. [1]

[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=Stee...