Quantization is widely adopted to reduce the computational cost of large language models (LLMs); however, its implications for fairness and safety, particularly in dynamic quantization and multilingual contexts, remain underexplored. In this work, we conduct a systematic study of how static and dynamic quantization methods impact fairness and safety across benchmarks measuring intrinsic and extrinsic bias and safety alignment. For fairness, we evaluate English, French, Dutch, Spanish, and Turkish; for safety, we focus on English, Korean, and Arabic. Our findings reveal that quantization consistently degrades fairness and safety, with dynamic methods demonstrating greater stability than static ones. Moreover, fairness degradation varies across languages, while safety deterioration is especially pronounced in non-English settings. To address these risks, we introduce Critical Weight Protection, a novel technique that identifies and preserves fairness- and safety-critical weights during quantization. This approach mitigates bias and safety deterioration without costly retraining or alignment, maintaining trustworthiness while retaining efficiency.

There are inconsistent conclusions about the impact of calibration data on Post-Training Quantization (PTQ) across studies, specifically regarding the factors that make the method effective, such as language that we explore in this work. We investigate this question by evaluating GPTQ and its mixed-precision variants across three low-bit settings, using language-targeted calibration sets across five language families. By evaluating performance across both perplexity and downstream tasks, we find that selecting an appropriate calibration language is generally beneficial, though model sensitivity varies with bit-width, target language, and task difficulty. We further show that mixed-precision techniques expand the representational capacity of low-bit models, occasionally allowing 2-bit models to approach 3-bit performance when paired with language-targeted calibration, while increased effective-bits remains small. Furthermore, we explore the correlation between these factors and observe that the calibration set sensitivity is influenced by different combinations of factors for each controlled variable, which can be explained by the downstream task results. Finally, while the benefit of language-targeted calibration transfers to sibling languages within the same family, the magnitude of this transfer does not reliably predicted by their closeness. Together, these findings show calibration data selection as a deliberate design choice in multilingual PTQ rather than an incidental one. Our code and result data can be accessed in https://github.com/rafifpriyo/multilingual-calibration-study.

Activation steering offers a lightweight mechanism for controlling model behavior, but existing methods depend on sentence-level contrastive data that is unavailable for most of the world’s languages. We show that multilingual sentiment lexicons are a viable substitute across three model families: aggregated multilingual lexicons outperform English-derived steering vectors for mid- and low-resource languages, while monolingual lexicons remain stronger for languages a model has specialized in during pre-training. The steering efficacy is mechanistically dictated by both representational density and pre-training priors. Large-magnitude English vectors overwhelm the compact representations of low-resource languages, causing "linguistic collapse." This fragility is mitigated by the norm-dampening of multilingual aggregation. Architectural biases, including bilingual specialization penalty and latent numerical rigidity, also heavily constrain cross-lingual steering responsiveness. Ultimately, our work establishes lexicons as a practical steering signal and offers a comprehensive mechanistic account of cross-lingual steering dynamics.

Understanding the multilingual mechanisms of large language models (LLMs) provides insight into how they process different languages, yet this remains challenging. Existing studies often focus on individual neurons, but their polysemantic nature makes it difficult to isolate language-specific units from cross-lingual representations. To address this, we explore sparse autoencoders (SAEs) for their ability to learn monosemantic features that represent concrete and abstract concepts across languages in LLMs. While some of these features are language-independent, the presence of language-specific features remains underexplored. In this work, we introduce SAE-LAPE, a method based on feature activation probability, to identify language-specific features within the feed-forward network. We find that many such features predominantly appear in the middle to late layers of the model and are interpretable. These features influence the model’s multilingual performance and language output, and can be used for language identification with performance comparable to fastText, along with more interpretability. Our code and complete figures are available [here](https://github.com/LyzanderAndrylie/language-specific-features).

This paper presents our approach for SemEval 2025 Task 11 Track A, focusing on multilabel emotion classification across 28 languages. We explore two main strategies: fully fine-tuning transformer models and classifier-only training, evaluating different settings such as fine-tuning strategies, model architectures, loss functions, encoders, and classifiers. Our findings suggest that training a classifier on top of prompt-based encoders such as mE5 and BGE yields significantly better results than fully fine-tuning XLMR and mBERT. Our best-performing model on the final leaderboard is an ensemble combining multiple BGE models, where CatBoost serves as the classifier, with different configurations. This ensemble achieves an average F1-macro score of 56.58 across all languages.

Language-specific neurons, units in LLMs that strongly correlate with individual languages have been shown to influence model behavior by deactivating them. However, their role in amplification remains underexplored.This work investigates the effect of amplifying language-specific neurons through interventions across 18 languages, including low-resource ones, using three models primarily trained in different languages. We compare amplification factors by their effectiveness in steering to the target language using a proposed Language Steering Shift (LSS) evaluation score, then evaluate it on downstream tasks: commonsense reasoning (XCOPA, XWinograd), knowledge (Include), and translation (FLORES). The optimal amplification steering factors effectively steer output toward nearly all tested languages. Intervention using this factor on downstream tasks improves self-language performance in some cases but generally degrades cross-language results. These findings highlight the effect of language-specific neurons in multilingual behavior, where amplification can be beneficial especially for low-resource languages, but provides limited advantage for cross-lingual transfer.

The performance of large language models (LLMs) continues to improve, as reflected in rising scores on standard benchmarks. However, the lack of transparency around training data raises concerns about potential overlap with evaluation sets and the fairness of reported results. Although prior work has proposed methods for detecting data leakage, these approaches primarily focus on identifying outliers and have not been evaluated under controlled simulated leakage conditions. In this work, we compare existing leakage detection techniques, namely permutation and n-gram-based methods, under a continual pretraining setup that simulates real-world leakage scenarios, and additionally explore a lightweight method we call semi-half question. We further introduce two efficient extensions, permutation-R and permutation-Q. While semi-half offers a low-cost alternative, our analysis shows that the n-gram method consistently achieves the highest F1-Score, performing competitively with permutation-Q. We also refine these techniques to support instance-level detection and reduce computational overhead. Leveraging the best-performing method, we create cleaned versions of MMLU and HellaSwag, and re-evaluate several LLMs. Our findings present a practical path toward more reliable and transparent evaluations, and we recommend contamination checks as a standard practice before releasing benchmark results.

Although region-specific large language models (LLMs) are increasingly developed, their safety remains underexplored, particularly in culturally diverse settings like Indonesia, where sensitivity to local norms is essential and highly valued by the community. In this work, we present IndoSafety, the first high-quality, human-verified safety evaluation dataset tailored for the Indonesian context, covering five language varieties: formal and colloquial Indonesian, along with three major local languages: Javanese, Sundanese, and Minangkabau. IndoSafety is constructed by extending prior safety frameworks to develop a taxonomy that captures Indonesia’s sociocultural context. We find that existing Indonesian-centric LLMs often generate unsafe outputs, particularly in colloquial and local language settings, while fine-tuning on IndoSafety significantly improves safety while preserving task performance. Our work highlights the critical need for culturally grounded safety evaluation and provides a concrete step toward responsible LLM deployment in multilingual settings. Warning: This paper contains example data that may be offensive, harmful, or biased.

Most Semantic Role Labeling (SRL) approaches are supervised methods which require a significant amount of annotated corpus, and the annotation requires linguistic expertise. In this paper, we propose a Multi-Task Active Learning framework for Semantic Role Labeling with Entity Recognition (ER) as the auxiliary task to alleviate the need for extensive data and use additional information from ER to help SRL. We evaluate our approach on Indonesian conversational dataset. Our experiments show that multi-task active learning can outperform single-task active learning method and standard multi-task learning. According to our results, active learning is more efficient by using 12% less of training data compared to passive learning in both single-task and multi-task setting. We also introduce a new dataset for SRL in Indonesian conversational domain to encourage further research in this area.

We propose keyphrases extraction technique to extract important terms from the healthcare user-generated contents. We employ deep learning architecture, i.e. Long Short-Term Memory, and leverage word embeddings, medical concepts from a knowledge base, and linguistic components as our features. The proposed model achieves 61.37% F-1 score. Experimental results indicate that our proposed approach outperforms the baseline methods, i.e. RAKE and CRF, on the task of extracting keyphrases from Indonesian health forum posts.