Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Cao, Xinyuan, Vempala, Santosh S.
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2512.07818
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866917133131186176
author Cao, Xinyuan
Vempala, Santosh S.
author_facet Cao, Xinyuan
Vempala, Santosh S.
contents Why do modern language models, trained to do well on next-word prediction, appear to generate coherent documents and capture long-range structure? Here we show that next-token prediction is provably powerful for learning longer-range structure, even with common neural network architectures. Specifically, we prove that optimizing next-token prediction over a Recurrent Neural Network (RNN) yields a model that closely approximates the training distribution: for held-out documents sampled from the training distribution, no algorithm of bounded description length limited to examining the next $k$ tokens, for any $k$, can distinguish between $k$ consecutive tokens of such documents and $k$ tokens generated by the learned language model following the same prefix. We provide polynomial bounds (in $k$, independent of the document length) on the model size needed to achieve such $k$-token indistinguishability, offering a complexity-theoretic explanation for the long-range coherence observed in practice.
format Preprint
id arxiv_https___arxiv_org_abs_2512_07818
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Provable Long-Range Benefits of Next-Token Prediction
Cao, Xinyuan
Vempala, Santosh S.
Machine Learning
Artificial Intelligence
Why do modern language models, trained to do well on next-word prediction, appear to generate coherent documents and capture long-range structure? Here we show that next-token prediction is provably powerful for learning longer-range structure, even with common neural network architectures. Specifically, we prove that optimizing next-token prediction over a Recurrent Neural Network (RNN) yields a model that closely approximates the training distribution: for held-out documents sampled from the training distribution, no algorithm of bounded description length limited to examining the next $k$ tokens, for any $k$, can distinguish between $k$ consecutive tokens of such documents and $k$ tokens generated by the learned language model following the same prefix. We provide polynomial bounds (in $k$, independent of the document length) on the model size needed to achieve such $k$-token indistinguishability, offering a complexity-theoretic explanation for the long-range coherence observed in practice.
title Provable Long-Range Benefits of Next-Token Prediction
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2512.07818