Adaptive multi-token prediction speeds up language model decoding
Adaptive Multi-Token Prediction: How far should we go?
Artificial IntelligenceMachine Learning
Summary
Language models need to guess the next words when creating sentences, which can be slow if done one word at a time. The authors study how to predict multiple words together to speed this up but adjust how many words they guess based on the context to avoid mistakes. They introduce a method called Adaptive Multi-Token Prediction (AMTP) that learns on the fly how far it can safely predict ahead without needing extra training. This approach makes language generation faster without losing accuracy, tested on several popular models and benchmarks.
What this means in practice
- •For natural language processing engineers: Implement adaptive prediction horizons to accelerate transformer-based language generation in production systems.
- •For machine learning infrastructure teams: Optimize model serving pipelines by integrating adaptive multi-token prediction to reduce inference latency with existing verifier models.
Authors
Emile Anand, Mohammed Rehan Parwani, Xinyuan Cao
Abstract
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 via acceptance survival, the probability that a draft prefix is fully accepted. Before each round, AMTP predicts how far the draft is likely to survive and picks the horizon with the highest expected tokens per second, learning freely from every verification pass. Theoretically, we show that any online learner's prediction error translates into throughput regret against the best horizon for each context, and in turn into a bound on wall-clock completion time. With online gradient descent, this regret is sublinear. Across five draft-target pairings built on Llama-3.1-8B-Instruct with Medusa heads, Meta's MTP models, and Qwen3-4B-Inst-2507-MTP, evaluated on AIME 2025 and GSM8K, AMTP decodes up to 4.1x faster than standard decoding and matches or outperforms the fixed window in 8 of 10 settings, with throughput gains of up to 10%.