Report
SwitchSD uses a model’s hidden states to choose between drafting and copying
A post introducing SwitchSD says the speculative-decoding method uses the target model’s hidden states to choose between neural drafting and copying from context. Its author says the work was accepted to NeurIPS 2026.
TLDR
SwitchSD’s author says the method uses a target language model’s hidden states to decide when to use a neural drafter and when to copy from context. The author says the work was accepted to NeurIPS 2026.
Combined views
154
1 Source, first seen 14h ago
7 reposts
