Saved in:
Bibliographic Details
Main Authors: Antoine, Elie, Béchet, Frédéric, Langlais, Philippe
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2412.16971
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866929644543934464
author Antoine, Elie
Béchet, Frédéric
Langlais, Philippe
author_facet Antoine, Elie
Béchet, Frédéric
Langlais, Philippe
contents This study investigates the behavior of model-integrated routers in Mixture of Experts (MoE) models, focusing on how tokens are routed based on their linguistic features, specifically Part-of-Speech (POS) tags. The goal is to explore across different MoE architectures whether experts specialize in processing tokens with similar linguistic traits. By analyzing token trajectories across experts and layers, we aim to uncover how MoE models handle linguistic information. Findings from six popular MoE models reveal expert specialization for specific POS categories, with routing paths showing high predictive accuracy for POS, highlighting the value of routing paths in characterizing tokens.
format Preprint
id arxiv_https___arxiv_org_abs_2412_16971
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Part-Of-Speech Sensitivity of Routers in Mixture of Experts Models
Antoine, Elie
Béchet, Frédéric
Langlais, Philippe
Computation and Language
This study investigates the behavior of model-integrated routers in Mixture of Experts (MoE) models, focusing on how tokens are routed based on their linguistic features, specifically Part-of-Speech (POS) tags. The goal is to explore across different MoE architectures whether experts specialize in processing tokens with similar linguistic traits. By analyzing token trajectories across experts and layers, we aim to uncover how MoE models handle linguistic information. Findings from six popular MoE models reveal expert specialization for specific POS categories, with routing paths showing high predictive accuracy for POS, highlighting the value of routing paths in characterizing tokens.
title Part-Of-Speech Sensitivity of Routers in Mixture of Experts Models
topic Computation and Language
url https://arxiv.org/abs/2412.16971