SpectR: Dynamically Composing LM Experts with Spectral Routing

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Fleshman, William, Van Durme, Benjamin
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911106511929344
author Fleshman, William
Van Durme, Benjamin
author_facet Fleshman, William
Van Durme, Benjamin
contents Training large, general-purpose language models poses significant challenges. The growing availability of specialized expert models, fine-tuned from pretrained models for specific tasks or domains, offers a promising alternative. Leveraging the potential of these existing expert models in real-world applications requires effective methods to select or merge the models best suited for a given task. This paper introduces SPECTR, an approach for dynamically composing expert models at each time step during inference. Notably, our method requires no additional training and enables flexible, token- and layer-wise model combinations. Our experimental results demonstrate that SPECTR improves routing accuracy over alternative training-free methods, increasing task performance across expert domains.
format Preprint
id arxiv_https___arxiv_org_abs_2504_03454
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle SpectR: Dynamically Composing LM Experts with Spectral Routing
Fleshman, William
Van Durme, Benjamin
Computation and Language
Artificial Intelligence
Machine Learning
Training large, general-purpose language models poses significant challenges. The growing availability of specialized expert models, fine-tuned from pretrained models for specific tasks or domains, offers a promising alternative. Leveraging the potential of these existing expert models in real-world applications requires effective methods to select or merge the models best suited for a given task. This paper introduces SPECTR, an approach for dynamically composing expert models at each time step during inference. Notably, our method requires no additional training and enables flexible, token- and layer-wise model combinations. Our experimental results demonstrate that SPECTR improves routing accuracy over alternative training-free methods, increasing task performance across expert domains.
title SpectR: Dynamically Composing LM Experts with Spectral Routing
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2504.03454