Paths Not Taken: Understanding and Mending the Multilingual Factual Recall Pipeline

TL;DR

This paper investigates why multilingual large language models often produce inconsistent factual answers across languages, identifies key causes of errors, and proposes interventions that significantly improve factual recall in low-resource languages.

Contribution

The study uncovers the internal pipeline of multilingual LLMs, identifies sources of factual inconsistency, and introduces language-agnostic interventions that enhance multilingual factual recall.

Findings

Over 35% increase in recall accuracy for low-resource languages

Identification of English-centric recall mechanism as a key factor

Effective vector interventions improve multilingual factual consistency

Abstract

Multilingual large language models (LLMs) often exhibit factual inconsistencies across languages, with significantly better performance in factual recall tasks in English than in other languages. The causes of these failures, however, remain poorly understood. Using mechanistic analysis techniques, we uncover the underlying pipeline that LLMs employ, which involves using the English-centric factual recall mechanism to process multilingual queries and then translating English answers back into the target language. We identify two primary sources of error: insufficient engagement of the reliable English-centric mechanism for factual recall, and incorrect translation from English back into the target language for the final answer. To address these vulnerabilities, we introduce two vector interventions, both independent of languages and datasets, to redirect the model toward better internal paths for higher factual consistency. Our interventions combined increase the recall accuracy by over 35 percent for the lowest-performing language. Our findings demonstrate how mechanistic insights can be used to unlock latent multilingual capabilities in LLMs.