Lucas Resck

Chain-of-Thought (CoT) faithfulness matters for monitoring, explaining and safely deploying large language models, yet it has been studied mostly in English. Languages differ in level, the amount of text available to learn from, so models know less in lower-resource languages, which may change how much they rely on their reasoning. To our knowledge, we are the first to relate CoT faithfulness to both resource level and task type (knowledge- or reasoning-intensive). On parallel multiple-choice questions from Multilingual Massive Multitask Language Understanding (MMMLU) in eight languages, faithfulness rises as resource level falls on knowledge-intensive tasks but not on reasoning-intensive ones, a gap that points the same way across four models and several faithfulness measures. We propose, and find support for, the Reliance Hypothesis: models rely more on their reasoning when they cannot recall the answer, as in lower-resource languages on knowledge-intensive tasks. Model confidence partly explains the effect of resource level, but not the knowledge-reasoning gap. Languages in which models are less confident rely more on reasoning, yet less confident answers within a language do not, a Simpson’s paradox. In lower-resource languages, models rely on their reasoning for more answers, not more strongly for each. CoT monitoring may therefore be most informative in lower-resource languages, where it is least often validated.

Multilingual large language models (MLLMs) often answer culturally ambiguous questions inconsistently across languages, even for a fixed user persona – e.g., Shakespeare in English but Cervantes in Spanish for the same British user. To robustly quantify this Cross-lingual Cultural Inconsistency, we introduce Singleton Fleiss’s Kappa, a metric mathematically resilient to hallucinations. For mitigation, we propose Cross-lingual Cultural Consistent Preference Optimisation (C-3PO), a consensus-driven alignment framework. C-3PO raises Kappa by 0.07–0.10 over unaligned models, above every prompting and representation steering baseline averaged across personas, whilst largely preserving explicit user identities, cultural neutrality and intrinsic cultural knowledge; where it does not, a soft-consensus variant recovers most of the cost. Although trained only on multiple-choice questions, C-3PO also improves consistency in open-ended generation. Empirical evaluations demonstrate this inconsistency is more pronounced for lower-resource languages like Indonesian and Persian. Finally, early decoding of intermediate layers reveals that MLLMs implicitly personalise outputs towards the prompt language’s stereotypical culture as forward-pass representations stabilise.

LLM evaluators (trained reward models and prompted LLM-as-a-Judge) are routinely validated via pairwise accuracy. In a multilingual setting, this operates under the premise that high pairwise accuracy implies reliable, language-neutral scoring. We show that this assumption does not hold across languages. We conduct experiments with semantically identical instruction–response pairs across 23 languages, and find that multilingual evaluators assign significantly different scores to different evaluation languages. We show that this bias is statistically significant and consistent across eight evaluators of different architectures and training paradigms, and is strongly correlated with language resource level: lower-resource languages are scored more generously. Meanwhile, these biases are invisible to pairwise accuracy: evaluators would achieve above 90% pairwise accuracy, yet have 43% difference in acceptance-rate across languages under a global decision threshold, meaning, for instance, that harmful content in lower-resource languages is more likely to pass safety filters. We then investigate why lower-resource languages receive higher rather than lower scores, and we find that model uncertainty (measured by negative log-likelihood) is linked with the effect: models tend to give higher scores when less confident; however, language identity remains a significant predictor after controlling for uncertainty, and the bias cannot be explained away by content difficulty alone, but is a structural, language-level misalignment.

Multilingual large language models (MLLMs) demonstrate state-of-the-art capabilities across diverse cross-lingual and multilingual tasks. Their complex internal mechanisms, however, often lack transparency, posing significant challenges in elucidating their internal processing of multilingualism, cross-lingual transfer dynamics and handling of language-specific features. This paper addresses this critical gap by presenting a survey of current explainability and interpretability methods specifically for MLLMs. To our knowledge, it is the first comprehensive review of its kind. Existing literature is categorised according to the explainability techniques employed, the multilingual tasks addressed, the languages investigated and available resources. The survey further identifies key challenges, distils core findings and outlines promising avenues for future research within this rapidly evolving domain.

Saliency post-hoc explainability methods are important tools for understanding increasingly complex NLP models. While these methods can reflect the model’s reasoning, they may not align with human intuition, making the explanations not plausible. In this work, we present a methodology for incorporating rationales, which are text annotations explaining human decisions, into text classification models. This incorporation enhances the plausibility of post-hoc explanations while preserving their faithfulness. Our approach is agnostic to model architectures and explainability methods. We introduce the rationales during model training by augmenting the standard cross-entropy loss with a novel loss function inspired by contrastive learning. By leveraging a multi-objective optimization algorithm, we explore the trade-off between the two loss functions and generate a Pareto-optimal frontier of models that balance performance and plausibility. Through extensive experiments involving diverse models, datasets, and explainability methods, we demonstrate that our approach significantly enhances the quality of model explanations without causing substantial (sometimes negligible) degradation in the original model’s performance.