Probing the Emergence of Cross-lingual Alignment during LLM Training

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11320%2F25%3ADUBYWBKX" target="_blank" >RIV/00216208:11320/25:DUBYWBKX - isvavai.cz</a>

Výsledek na webu

<a href="https://www.scopus.com/inward/record.uri?eid=2-s2.0-85205292172&partnerID=40&md5=ad80761ce6b51e896ec97a724df353c6" target="_blank" >https://www.scopus.com/inward/record.uri?eid=2-s2.0-85205292172&partnerID=40&md5=ad80761ce6b51e896ec97a724df353c6</a>

DOI - Digital Object Identifier

—

Jazyk výsledku

angličtina

Název v původním jazyce

Probing the Emergence of Cross-lingual Alignment during LLM Training

Popis výsledku v původním jazyce

Multilingual Large Language Models (LLMs) achieve remarkable levels of zero-shot cross-lingual transfer performance. We speculate that this is predicated on their ability to align languages without explicit supervision from parallel sentences. While representations of translationally equivalent sentences in different languages are known to be similar after convergence, however, it remains unclear how such cross-lingual alignment emerges during pretraining of LLMs. Our study leverages intrinsic probing techniques, which identify which subsets of neurons encode linguistic features, to correlate the degree of cross-lingual neuron overlap with the zero-shot cross-lingual transfer performance for a given model. In particular, we rely on checkpoints of BLOOM, a multilingual autoregressive LLM, across different training steps and model scales. We observe a high correlation between neuron overlap and downstream performance, which supports our hypothesis on the conditions leading to effective cross-lingual transfer. Interestingly, we also detect a degradation of both implicit alignment and multilingual abilities in certain phases of the pre-training process, providing new insights into the multilingual pretraining dynamics. © 2024 Association for Computational Linguistics.

Název v anglickém jazyce

Probing the Emergence of Cross-lingual Alignment during LLM Training

Popis výsledku anglicky

Multilingual Large Language Models (LLMs) achieve remarkable levels of zero-shot cross-lingual transfer performance. We speculate that this is predicated on their ability to align languages without explicit supervision from parallel sentences. While representations of translationally equivalent sentences in different languages are known to be similar after convergence, however, it remains unclear how such cross-lingual alignment emerges during pretraining of LLMs. Our study leverages intrinsic probing techniques, which identify which subsets of neurons encode linguistic features, to correlate the degree of cross-lingual neuron overlap with the zero-shot cross-lingual transfer performance for a given model. In particular, we rely on checkpoints of BLOOM, a multilingual autoregressive LLM, across different training steps and model scales. We observe a high correlation between neuron overlap and downstream performance, which supports our hypothesis on the conditions leading to effective cross-lingual transfer. Interestingly, we also detect a degradation of both implicit alignment and multilingual abilities in certain phases of the pre-training process, providing new insights into the multilingual pretraining dynamics. © 2024 Association for Computational Linguistics.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

—

Návaznosti

—

Rok uplatnění

2024

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

Proc. Annu. Meet. Assoc. Comput Linguist.

ISBN

979-889176099-8

ISSN

0736-587X

e-ISSN

—

Počet stran výsledku

15

Strana od-do

12159-12173

Název nakladatele

Association for Computational Linguistics (ACL)

Místo vydání

—

Místo konání akce

Hybrid, Bangkok

Datum konání akce

1. 1. 2025

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—