Cross-Lingual Activation Steering for Multilingual Language Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Pokharel, Rhitabrat, Agrawal, Ameeta, Nagar, Tanay
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908783304769536
author Pokharel, Rhitabrat
Agrawal, Ameeta
Nagar, Tanay
author_facet Pokharel, Rhitabrat
Agrawal, Ameeta
Nagar, Tanay
contents Large language models exhibit strong multilingual capabilities, yet significant performance gaps persist between dominant and non-dominant languages. Prior work attributes this gap to imbalances between shared and language-specific neurons in multilingual representations. We propose Cross-Lingual Activation Steering (CLAS), a training-free inference-time intervention that selectively modulates neuron activations. We evaluate CLAS on classification and generation benchmarks, achieving average improvements of 2.3% (Acc.) and 3.4% (F1) respectively, while maintaining high-resource language performance. We discover that effective transfer operates through functional divergence rather than strict alignment; performance gains correlate with increased language cluster separation. Our results demonstrate that targeted activation steering can unlock latent multilingual capacity in existing models without modification to model weights.
format Preprint
id arxiv_https___arxiv_org_abs_2601_16390
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Cross-Lingual Activation Steering for Multilingual Language Models
Pokharel, Rhitabrat
Agrawal, Ameeta
Nagar, Tanay
Computation and Language
Artificial Intelligence
Large language models exhibit strong multilingual capabilities, yet significant performance gaps persist between dominant and non-dominant languages. Prior work attributes this gap to imbalances between shared and language-specific neurons in multilingual representations. We propose Cross-Lingual Activation Steering (CLAS), a training-free inference-time intervention that selectively modulates neuron activations. We evaluate CLAS on classification and generation benchmarks, achieving average improvements of 2.3% (Acc.) and 3.4% (F1) respectively, while maintaining high-resource language performance. We discover that effective transfer operates through functional divergence rather than strict alignment; performance gains correlate with increased language cluster separation. Our results demonstrate that targeted activation steering can unlock latent multilingual capacity in existing models without modification to model weights.
title Cross-Lingual Activation Steering for Multilingual Language Models
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2601.16390