Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Wu, Wilson, Morris, John X., Levine, Lionel
Format: Preprint
Veröffentlicht: 2024
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2404.00859
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866914896781770752
author Wu, Wilson
Morris, John X.
Levine, Lionel
author_facet Wu, Wilson
Morris, John X.
Levine, Lionel
contents Do transformers "think ahead" during inference at a given position? It is known transformers prepare information in the hidden states of the forward pass at time step $t$ that is then used in future forward passes $t+τ$. We posit two explanations for this phenomenon: pre-caching, in which off-diagonal gradient terms present during training result in the model computing features at $t$ irrelevant to the present inference task but useful for the future, and breadcrumbs, in which features most relevant to time step $t$ are already the same as those that would most benefit inference at time $t+τ$. We test these hypotheses by training language models without propagating gradients to past timesteps, a scheme we formalize as myopic training. In a constructed synthetic data setting, we find clear evidence for pre-caching. In the autoregressive language modeling setting, our experiments are more suggestive of the breadcrumbs hypothesis, though pre-caching increases with model scale.
format Preprint
id arxiv_https___arxiv_org_abs_2404_00859
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Do language models plan ahead for future tokens?
Wu, Wilson
Morris, John X.
Levine, Lionel
Machine Learning
Computation and Language
Do transformers "think ahead" during inference at a given position? It is known transformers prepare information in the hidden states of the forward pass at time step $t$ that is then used in future forward passes $t+τ$. We posit two explanations for this phenomenon: pre-caching, in which off-diagonal gradient terms present during training result in the model computing features at $t$ irrelevant to the present inference task but useful for the future, and breadcrumbs, in which features most relevant to time step $t$ are already the same as those that would most benefit inference at time $t+τ$. We test these hypotheses by training language models without propagating gradients to past timesteps, a scheme we formalize as myopic training. In a constructed synthetic data setting, we find clear evidence for pre-caching. In the autoregressive language modeling setting, our experiments are more suggestive of the breadcrumbs hypothesis, though pre-caching increases with model scale.
title Do language models plan ahead for future tokens?
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2404.00859