aiqbench analysis
This research addresses cross-lingual factual inconsistency in LLMs by comparing inference-time interventions like persona prompting against internal representation manipulation. The study finds that simple contextual steering offers a superior balance of efficacy and safety compared to invasive methods that risk knowledge degradation or lack transferability. These findings provide practical guidance for deploying robust multilingual models without relying on permanent weight modifications.
Inference-Time Steering for Cross-Lingual Factual Consistency in LLMs
arXiv:2607.19243v1 Announce Type: cross Abstract: Although Large Language Models (LLMs) demonstrate remarkable multilingual fluency, their internal knowledge representations remain disproportionately biased toward high-resource languages. This leads to cross-lingual factual inconsistency, where they shift their empirical answer distributions based solely on the prompt language. We investigate whether these biases can be mitigated at inference time, forcing an English-prompted model to answer as if it were queried in target languages (German, Spanish, Bulgarian), and evaluate four intervention strategies: zero-shot contextual steering (persona prompting), internal representation manipulation via Contrastive Activation Addition (CAA), and lightweight weight modification via Direct Preference Optimization (DPO) trained on benchmark-derived factual data as well as conceptual generalization data. To assess alignment, we curate a multilingual factual dataset alongside a novel generalization benchmark comprising culturally rooted queries to determine whether factual interventions transfer to broader target-centric preferences. Experiments on Gemma 3 12B Instruct reveal persona prompting to be the strongest overall intervention, balancing efficacy, safety, and out-of-domain generalization. While CAA yields sharp inconsistency benchmark shifts, it is configuration-sensitive and risks knowledge degradation. DPO-based adapters offer permanent, yet narrower and less transferable gains. These findings suggest that cross-lingual inconsistency is at least partly a selection problem, and that simple contextual interventions may outperform more invasive methods for robust, transferable alignment.
Abstract:Although Large Language Models (LLMs) demonstrate remarkable multilingual fluency, their internal knowledge representations remain disproportionately biased toward high-resource languages. This leads to cross-lingual factual inconsistency, where they shift their empirical answer distributions based solely on the prompt language. We investigate whether these biases can be mitigated at inference time, forcing an English-prompted model to answer as if it were queried in target languages (German, Spanish, Bulgarian), and evaluate four intervention strategies: zero-shot contextual steering (persona prompting), internal representation manipulation via Contrastive Activation Addition (CAA), and lightweight weight modification via Direct Preference Optimization (DPO) trained on benchmark-derived factual data as well as conceptual generalization data. To assess alignment, we curate a multilingual factual dataset alongside a novel generalization benchmark comprising culturally rooted queries to determine whether factual interventions transfer to broader target-centric preferences. Experiments on Gemma 3 12B Instruct reveal persona prompting to be the strongest overall intervention, balancing efficacy, safety, and out-of-domain generalization. While CAA yields sharp inconsistency benchmark shifts, it is configuration-sensitive and risks knowledge degradation. DPO-based adapters offer permanent, yet narrower and less transferable gains. These findings suggest that cross-lingual inconsistency is at least partly a selection problem, and that simple contextual interventions may outperform more invasive methods for robust, transferable alignment.
8 pages (21 in total), 2 figures, 4 tables. Original manuscript for a Guided Research project conducted at the Technical University of Munich, detailing the complete methodology, full data pipeline, and comprehensive experimental results. A related, condensed subset of this work was subsequently adapted and published at the StereACuLT 2026 workshop
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2607.19243 [cs.CL]
(or arXiv:2607.19243v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2607.19243 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 353 words
Related coverage
Cross-Lingual Representation Alignment by Token-Level Optimal Transport in a Language-Agnostic Space
arXiv · Sep 9, 2026
Latent-Space Intervention for Cross-Lingual Factual Consistency: Consistency Improvements without Accuracy Drops
arXiv · Sep 1, 2026
[AINews] The Future of Latent Space
Latent Space · Sep 25, 2026
The Continual Constitutional Convention: How language models revise the documents that govern them
AI Alignment Forum · Oct 13, 2026