Parallel Token Prediction for Language Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Draxler, Felix, Will, Justus, Sofian, Farrin Marouf, Karaletsos, Theofanis, Singh, Sameer, Mandt, Stephan
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910041367379968
author Draxler, Felix
Will, Justus
Sofian, Farrin Marouf
Karaletsos, Theofanis
Singh, Sameer
Mandt, Stephan
author_facet Draxler, Felix
Will, Justus
Sofian, Farrin Marouf
Karaletsos, Theofanis
Singh, Sameer
Mandt, Stephan
contents Autoregressive decoding in language models is inherently slow, generating only one token per forward pass. We propose Parallel Token Prediction (PTP), a general-purpose framework for predicting multiple tokens in a single model call. PTP moves the source of randomness from post-hoc sampling to random input variables, making future tokens deterministic functions of those inputs and thus jointly predictable in a single forward pass. We prove that a single PTP call can represent arbitrary dependencies between tokens. PTP is trained by distilling an existing model or through inverse autoregressive training without a teacher. Experimentally, PTP achieves a 2.4x speedup on a diverse-task speculative decoding benchmark. We provide code and checkpoints at https://github.com/mandt-lab/ptp.
format Preprint
id arxiv_https___arxiv_org_abs_2512_21323
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Parallel Token Prediction for Language Models
Draxler, Felix
Will, Justus
Sofian, Farrin Marouf
Karaletsos, Theofanis
Singh, Sameer
Mandt, Stephan
Computation and Language
Machine Learning
Autoregressive decoding in language models is inherently slow, generating only one token per forward pass. We propose Parallel Token Prediction (PTP), a general-purpose framework for predicting multiple tokens in a single model call. PTP moves the source of randomness from post-hoc sampling to random input variables, making future tokens deterministic functions of those inputs and thus jointly predictable in a single forward pass. We prove that a single PTP call can represent arbitrary dependencies between tokens. PTP is trained by distilling an existing model or through inverse autoregressive training without a teacher. Experimentally, PTP achieves a 2.4x speedup on a diverse-task speculative decoding benchmark. We provide code and checkpoints at https://github.com/mandt-lab/ptp.
title Parallel Token Prediction for Language Models
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2512.21323