ConceptImplicit Preference Pairs
Implicit Preference Pairs
In speculative decoding, contrastive data formed naturally by token sequences that the target model accepts and rejects, used as preference signals for reinforcement learning without manual annotation.
Timeline (last 90 days)
Oct 22
RCSD uses the acceptance and rejection sequences naturally produced during speculative inference to construct implicit preference pairs as preference signals in reinforcement learning to optimize the draft model.
Unverified50%