Tag
1 articles
SpecKV adapts speculative decoding’s token budget per step, using draft-model signals to beat fixed gamma across compression settings.