Toward Consistent World Models with Multi-Token Prediction and Latent Semantic Enhancement

TL;DR

This paper investigates how Multi-Token Prediction (MTP) can improve internal world models in Large Language Models, introduces a new method LSE-MTP to reduce hallucinations, and demonstrates its effectiveness on synthetic and real data.

Contribution

It provides a theoretical analysis of MTP's bias, identifies structural hallucinations, and proposes LSE-MTP to enhance state representation alignment and robustness.

Findings

LSE-MTP reduces structural hallucinations in models.

MTP promotes convergence toward internal belief states.

LSE-MTP improves robustness to perturbations.

Abstract

Whether Large Language Models (LLMs) develop coherent internal world models remains a core debate. While conventional Next-Token Prediction (NTP) focuses on one-step-ahead supervision, Multi-Token Prediction (MTP) has shown promise in learning more structured representations. In this work, we provide a theoretical perspective analyzing the gradient inductive bias of MTP, supported by empirical evidence, showing that MTP promotes the convergence toward internal belief states by inducing representational contractivity via gradient coupling. However, we reveal that standard MTP often suffers from structural hallucinations, where discrete token supervision encourages illegal shortcuts in latent space that violate environmental constraints. To address this, we propose a novel method Latent Semantic Enhancement MTP (LSE-MTP), which anchors predictions to ground-truth hidden state trajectories. Experiments on synthetic graphs and real-world Manhattan Taxi Ride show that LSE-MTP effectively bridges the gap between discrete tokens and continuous state representations, enhancing representation alignment, reducing structural hallucinations, and improving robustness to perturbations.