How Transformers Learn to Plan via Multi-Token Prediction
For language model training, MTP offers a principled alternative to NTP that inherently biases optimization toward robust planning circuits, with both empirical and theoretical validation.
Multi-token prediction (MTP) outperforms next-token prediction (NTP) on planning tasks, achieving up to 20% higher accuracy on synthetic graph path-finding and strong gains on Countdown and SAT benchmarks. Theoretically, MTP induces a reverse reasoning process via gradient decoupling.
While next-token prediction (NTP) has been the standard objective for training language models, it often struggles to capture global structure in reasoning tasks. Multi-token prediction (MTP) has recently emerged as a promising alternative, yet its underlying mechanisms remain poorly understood. In this paper, we study how MTP facilitates reasoning, with a focus on planning. Empirically, we show that MTP consistently outperforms NTP on both synthetic graph path-finding tasks and more realistic reasoning benchmarks, such as Countdown and boolean satisfiability problems. Theoretically, we analyze a simplified two-layer Transformer on a star graph task. We prove that MTP induces a two-stage reverse reasoning process: the model first attends to the end node and then reconstructs the path by tracing intermediate nodes backward. This behavior arises from a gradient decoupling property of MTP, which provides a cleaner training signal compared to NTP. Ultimately, our results highlight how multi-token objectives inherently bias optimization toward robust and interpretable reasoning circuits.