# Inference-Time Steering for Cross-Lingual Factual Consistency in LLMs

arXiv: https://arxiv.org/abs/2607.19243 · HTML: https://mlsanity.com/paper/2607.19243/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in LLM preference optimization methods, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 15 of its 24 references that are in the map.
- First posted on arXiv 21 Jul 2026.

## Placed in

- LLM alignment / RLHF / preference optimisation / [LLM preference optimization methods](https://mlsanity.com/alignment/llm-preference-optimization-methods/)

## Builds on (most cited first, 15 of 15)

- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Representation Engineering: A Top-Down Approach to AI Transparency](https://mlsanity.com/paper/2310.01405/) (2023)
- [Steering Llama 2 via Contrastive Activation Addition](https://mlsanity.com/paper/2312.06681/) (2023)
- [Steering Language Models With Activation Engineering](https://mlsanity.com/paper/2308.10248/) (2023)
- [Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive](https://mlsanity.com/paper/2402.13228/) (2024)
- [The Semantic Hub Hypothesis: Language Models Share Semantic Representations Across Languages and Modalities](https://mlsanity.com/paper/2411.04986/) (2024)
- [Lost in Multilinguality: Dissecting Cross-lingual Factual Inconsistency in Transformer Language Models](https://mlsanity.com/paper/2504.04264/) (2025)
- [The Prompt Makes the Person(a): A Systematic Evaluation of Sociodemographic Persona Prompting for Large Language Models](https://mlsanity.com/paper/2507.16076/) (2025)
- [CulturalBench: A Robust, Diverse, and Challenging Cultural Benchmark by Human-AI CulturalTeaming](https://mlsanity.com/paper/2410.02677/) (2024)
- [Multilingual Prompting for Improving LLM Generation Diversity](https://mlsanity.com/paper/2505.15229/) (2025)
- [Multi-FAct: Assessing Factuality of Multilingual LLMs using FActScore](https://mlsanity.com/paper/2402.18045/) (2024)
- [Aligning LLMs for Multilingual Consistency in Enterprise Applications](https://mlsanity.com/paper/2509.23659/) (2025)
- [From Unaligned to Aligned: Scaling Multilingual LLMs with Multi-Way Parallel Corpora](https://mlsanity.com/paper/2505.14045/) (2025)
- [AlignX: Advancing Multilingual Large Language Models with Multilingual Representation Alignment](https://mlsanity.com/paper/2509.24338/) (2025)
- [Selected Languages are All You Need for Cross-lingual Truthfulness Transfer](https://mlsanity.com/paper/2406.14434/) (2024)

## Built on by (most cited first, 1 of 1)

- [Why Pretraining Fails to Share Cross-Lingual Knowledge](https://mlsanity.com/paper/2609.19291/) (2026)

## Abstract (arXiv)

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.