Exit-Guided Speculative Decoding via TEV and ExitTrain

Tree Exit Verification (TEV) introduces an exact, level-parallel procedure for speculative decoding that uses a single exit-node decision to optimize target coverage. The framework includes ExitTrain, which uses node-level feedback to train draft trees to better match target model probability distributions.