Efficient Training-Free Multi-Token Prediction via Embedding-Space Probing

2026-03-18

AI Summary

This work proposes a training-free, auxiliary-model-free approach for multi-token parallel prediction to enhance the inference efficiency of large language models. By dynamically constructing masked token probes in the embedding space, the method activates the model’s inherent multi-step prediction capability. It integrates top-K candidate tree construction, lightweight pruning, and a parallel verification decoding mechanism to achieve efficient and lossless generation. Notably, this is the first method to introduce a training-free embedding-space probing mechanism that leverages only the original model architecture to accurately predict multi-token sequences. Experiments demonstrate consistent improvements: on LLaMA3 and Qwen3, it increases accepted sequence length by approximately 12% and 8–12%, respectively, and boosts throughput by 15–19%, significantly outperforming existing training-free baselines.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Large language models (LLMs) exhibit latent multi-token prediction (MTP) capabilities despite being trained solely for next-token generation. We propose a simple, training-free MTP approach that probes an LLM using on-the-fly mask tokens drawn from its embedding space, enabling parallel prediction of future tokens without modifying model weights or relying on auxiliary draft models. Our method constructs a speculative token tree by sampling top-K candidates from mask-token logits and applies a lightweight pruning strategy to retain high-probability continuations. During decoding, candidate predictions are verified in parallel, resulting in lossless generation while substantially reducing the number of model calls and improving token throughput. Across benchmarks, our probing-based MTP consistently outperforms existing training-free baselines, increasing acceptance length by approximately 12\% on LLaMA3 and 8--12\% on Qwen3, and achieving throughput gains of up to 15--19\%. Finally, we provide theoretical insights and empirical evidence showing that decoder layers naturally align mask-token representations with next-token states, enabling accurate multi-step prediction without retraining or auxiliary models.

Problem

• multi-token prediction
• large language models
• training-free
• speculative decoding
• embedding-space probing

Innovation

• training-free
• multi-token prediction
• embedding-space probing
• speculative decoding
• mask token