Adaptive Multi-Token Prediction: How far should we go?
[Sztuczna inteligencja] Autorzy: Emile Anand, Mohammed Rehan Parwani, Xinyuan Cao
Multi-token prediction (MTP) can accelerate language-model decoding by drafting several tokens and verifying them in a single target-model pass. However, the best speculation horizon changes with context: longer drafts offer greater amortization when reliable, but can waste compute on suffixes discarded after a rejection. Unlike autoregressive drafting, where a stopping rule can decide token by token, MTP must fix the horizon before drafting, as all heads fire in the same forward pass. We therefore formulate horizon selection as an online control problem, using verification feedback to learn how far the draft can be trusted in each context, and introduce Adaptive Multi-Token Prediction (AMTP): a lightweight controller that requires no retraining and leaves the exact verifier unchanged. To guide horizon selection, we characterize the expected number of tokens committed at each horizon vi…
- Kategoria arxiv
- cs.AI
- Kategoria
- Sztuczna inteligencja