ParaScopes: What do Language Models Activations Encode About Future Text?

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Pochinkov, Nicky, Volkova, Yulia, Vasileva, Anna, Chereddy, Sai V R
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915589492047872
author Pochinkov, Nicky
Volkova, Yulia
Vasileva, Anna
Chereddy, Sai V R
author_facet Pochinkov, Nicky
Volkova, Yulia
Vasileva, Anna
Chereddy, Sai V R
contents Interpretability studies in language models often investigate forward-looking representations of activations. However, as language models become capable of doing ever longer time horizon tasks, methods for understanding activations often remain limited to testing specific concepts or tokens. We develop a framework of Residual Stream Decoders as a method of probing model activations for paragraph-scale and document-scale plans. We test several methods and find information can be decoded equivalent to 5+ tokens of future context in small models. These results lay the groundwork for better monitoring of language models and better understanding how they might encode longer-term planning information.
format Preprint
id arxiv_https___arxiv_org_abs_2511_00180
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle ParaScopes: What do Language Models Activations Encode About Future Text?
Pochinkov, Nicky
Volkova, Yulia
Vasileva, Anna
Chereddy, Sai V R
Computation and Language
Machine Learning
Interpretability studies in language models often investigate forward-looking representations of activations. However, as language models become capable of doing ever longer time horizon tasks, methods for understanding activations often remain limited to testing specific concepts or tokens. We develop a framework of Residual Stream Decoders as a method of probing model activations for paragraph-scale and document-scale plans. We test several methods and find information can be decoded equivalent to 5+ tokens of future context in small models. These results lay the groundwork for better monitoring of language models and better understanding how they might encode longer-term planning information.
title ParaScopes: What do Language Models Activations Encode About Future Text?
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2511.00180