Speculative Decoding

Self-speculative decoding that reformulates draft model selection as a knapsack problem to maximize tokens-per-time throughput, accounting for context-dependent attention overhead - ___[ICML 2026](http://icml.cc/Conferences/2026)___