Multilingual embedding models often exhibit uneven representational quality, heavily favoring high-resource languages like English. However, conventional retrieval systems that rely exclusively on source-language queries fail to exploit the superior semantic expressiveness of these high-resource subspaces. To address this, we propose Query-Synergy, a training-free approach to improving retrieval performance using multilingual embeddings. Our method utilizes additional queries in English to complement source language queries and integrates similarity scores from both queries, effectively enhancing retrieval performance. We evaluate our approach across five languages (Arabic, Chinese, Greek, Thai, and Turkish) using four multilingual embedding models on two datasets. Our experiments show that this approach outperforms conventional source query retrieval methods, achieving superior nDCG scores across various configurations and translation settings. These results confirm that Query-Synergy is a simple yet effective method for retrieval across multiple languages.

With the increasing utilization of multilingual text information, Cross-Lingual Information Retrieval (CLIR) has become a crucial research area. However, the impact of training data composition on CLIR and Mono-Lingual Information Retrieval (Mono-IR) performance remains underexplored. To investigate this data-centric aspect, we construct linguistically parallel Korean-English datasets and train multilingual retrieval models with various language combinations. Our experiments reveal that the language composition of training data significantly influence IR performance, exhibiting important inter-lingual correlations: Using specific language pairs improves CLIR performance, while declines Mono-IR performance. Our work demonstrates that simple weight-averaged model merging can effectively mitigate this trade-off, achieving strong CLIR results while preserving Mono-IR capabilities. Our findings highlight the effects of linguistic configuration of training data on both CLIR and Mono-IR, and present model merging as a viable strategy to optimize performance across these tasks.

Refusal on a safety benchmark does not reveal how stable that behavior will remain after model updates. Benign downstream fine-tuning can weaken refusal, yet behavioral evaluations typically expose this fragility only after an intervention. We introduce SKIN-DEEP, a geometric diagnostic that examines the unmodified model’s residual-stream activations. It compares aligned and base checkpoints to identify safety-separating directions, tests their behavioral relevance through ablation, and summarizes the layer-wise pattern in the Geometric Fragility Score (GFS). Across twenty-one instruction-tuned models, harmful requests and benign instructions exhibit a recurring low-rank separation pattern. Selected direction ablations weaken refusal, with the effective direction varying across models. In benign low-rank fine-tuning experiments, the initially safe model with the lowest score before fine-tuning has the lowest harmful-compliance rate when trained on the largest tested set of harmless examples. These findings connect representation geometry to subsequent behavioral susceptibility and support activation-based diagnostics as a complement to refusal tests. Our code is available at https://github.com/js-lee-AI/skin-deep.

Current cultural evaluations for large language models (LLMs) often reduce culture to single-turn factual recall via MCQs, failing to capture a common use case: users seeking practical help over multiple turns in culturally grounded scenarios. We introduce CultureConverse, a scalable, multilingual simulation and evaluation harness for culturally grounded assistant dialogue that covers 10 East and Southeast Asian regions, 58 subgroup identities, and 7 domains. Each simulated and evaluated episode produces a scored interaction where the assistant assists the user and infers cultural constraints from partial information. The resulting CultureConverse-DS dataset contains 14,610 benchmark (evaluation) episodes and 274,295 oracle-guided (gold-mode) dialogues. In our benchmark evaluation of 18 models, GPT-5 mini achieves the highest assistance quality. Human annotation experiments suggest that our evaluation framework is a sufficient proxy for human judgment. Performance gains from fine-tuning on 27,860 high-quality CultureConverse-DS samples improve in-domain assistance and transfer out-of-domain to cultural MCQ and safety classification benchmarks. We release the harness, both splits, and judge prompts to support interactive evaluation of cultural competency.

The alignment of Large Language Models heavily relies on English-centric high-quality preference data, which often leads to suboptimal performance in other languages. In this paper, we propose Cross-lingual Ranking Preference Optimization (CRPO), a novel framework that leverages robust preference knowledge from English to facilitate preference alignment in the target language. We design a hierarchical structure within parallel preference pairs across the target language and English to jointly optimize intra- and inter-lingual preferences, thereby enhancing language adaptation and output quality. Building on the LambdaLoss framework, CRPO goes beyond the binary comparison based optimization by providing a relative ranking signal across multiple candidate responses. Our experiments across five languages with varying resource scales demonstrate that CRPO consistently outperforms standard approaches in both instruction-following and knowledge utilization capability. Notably, the robust performance gains observed across various weighting schemes further validate the empirical effectiveness of our hierarchical design in a multilingual setup. Furthermore, our findings highlight that CRPO significantly improves both reward margins and the log-probability of desirable responses, contributing to a more stable preference manifold for cross-lingual alignment.

Existing multilingual embedding models often encounter challenges in cross-lingual scenarios due to imbalanced linguistic resources and less consideration of cross-lingual alignment during training. Although standardized contrastive learning approaches for cross-lingual adaptation are widely adopted, they may struggle to capture fundamental alignment between languages and degrade performance in well-aligned languages such as English. To address these challenges, we propose Cross-Lingual Enhancement in RetrievAl via Reverse-training (CLEAR), a novel loss function utilizing a reverse training scheme to improve retrieval performance across diverse cross-lingual retrieval scenarios. CLEAR leverages an English passage as a bridge to strengthen alignments between the target language and English, ensuring robust performance in the cross-lingual retrieval task. Our extensive experiments demonstrate that CLEAR achieves notable improvements in cross-lingual scenarios, with gains up to 15%, particularly in low-resource languages, while minimizing performance degradation in English. Furthermore, our findings highlight that CLEAR offers promising effectiveness even in multilingual training, suggesting its potential for broad application and scalability. We release the code at https://github.com/dltmddbs100/CLEAR.

Through numerous endeavors, large language models (LLMs) have witnessed significant advancements in their instruction-following capability. However, we discern that LLMs are prone to generate responses to instruction-formatted statements in an instinctive manner, rather than comprehending the underlying user intention reside within the given instructions. We also recognize that the significance of instruction understanding capability is largely overlooked in most of LLM evaluation benchmarks. To ensure more comprehensive evaluation on the instruction understanding capability of LLM, we propose Intention of Instruction (IntInst) benchmark, which primary objective is to distinguish the appropriate instruction that accurately instruct to generate a given context. IntInst presents four instruction candidates and requires LLMs to select one among them. Through extensive experiments with several instruction-tuned LLMs, we reveal that most LLMs struggle to grasp the actual intention concealed in the instruction and thoroughly analyze the factors influencing instruction understanding.

Recent advancements in Large Language Models (LLMs) have significantly enhanced interactions between users and models. These advancements concurrently underscore the need for rigorous safety evaluations due to the manifestation of social biases, which can lead to harmful societal impacts. Despite these concerns, existing benchmarks may overlook the intrinsic weaknesses of LLMs, which can generate biased responses even with simple adversarial instructions. To address this critical gap, we introduce a new benchmark, Fairness Benchmark in LLM under Extreme Scenarios (FLEX), designed to test whether LLMs can sustain fairness even when exposed to prompts constructed to induce bias. To thoroughly evaluate the robustness of LLMs, we integrate prompts that amplify potential biases into the fairness assessment. Comparative experiments between FLEX and existing benchmarks demonstrate that traditional evaluations may underestimate the inherent risks in models. This highlights the need for more stringent LLM evaluation benchmarks to guarantee safety and fairness.

Large Language Models (LLMs) are increasingly incorporating multilingual capabilities, fueling the demand to transfer them into target language-specific models. However, most approaches, which blend the source model’s embedding by replacing the source vocabulary with the target language-specific vocabulary, may constrain expressive capacity in the target language since the source model is predominantly trained on English data. In this paper, we propose Semantic Aware Linear Transfer (SALT), a novel cross-lingual transfer technique that recycles embeddings from target language Pre-trained Language Models (PLMs) to transmit the deep representational strengths of PLM-derived embedding to LLMs. SALT derives unique regression lines based on the similarity in the overlap of the source and target vocabularies to handle each non-overlapping token’s embedding space. Our extensive experiments show that SALT significantly outperforms other transfer methods, achieving lower loss and faster convergence during language adaptation. Notably, SALT achieves remarkable performance in cross-lingual understanding setups compared to other methods. Furthermore, we highlight the scalable use of PLMs to enhance the functionality of contemporary LLMs by conducting experiments with varying architectures.

The approaches that guide Large Language Models (LLMs) to emulate human reasoning during response generation have emerged as an effective method for enabling them to solve complex problems in a step-by-step manner, thereby achieving superior performance. However, most existing approaches using few-shot prompts to generate responses heavily depend on the provided examples, limiting the utilization of the model’s inherent reasoning capabilities. Moreover, constructing task-specific few-shot prompts is often costly and may lead to inconsistencies across different tasks. In this work, we introduce Template Oriented Reasoning (TORSO), which elicits the model to utilize internal reasoning abilities to generate proper responses across various tasks without the need for manually crafted few-shot examples. Our experimental results demonstrate that TORSO achieves strong performance on diverse LLMs benchmarks with reasonable rationales.

Large Language Models (LLMs) have rapidly advanced, with domain-specific expert models emerging to handle specialized tasks across various fields. However, the predominant focus on English-centric models demands extensive data, making it challenging to develop comparable models for middle and low-resource languages. To address this limitation, we introduce Migrate, a novel method that leverages open-source static embedding models and up to 3 million tokens of code-switching data to facilitate the seamless transfer of embeddings to target languages. Migrate enables effective cross-lingual adaptation without requiring large-scale domain-specific corpora in the target language, promoting the accessibility of expert LLMs to a diverse range of linguistic communities. Our experimental results demonstrate that Migrate significantly enhances model performance in target languages, outperforming baseline and existing cross-lingual transfer methods. This approach provides a practical and efficient solution for extending the capabilities of domain-specific expert models.

Large Language Models (LLMs) have significantly impacted various fields requiring advanced linguistic understanding, yet concerns regarding their inherent biases and ethical considerations have also increased. Notably, LLMs have been critiqued for perpetuating stereotypes against diverse groups based on race, sexual orientation, and other attributes. However, most research analyzing these biases has predominantly focused on communities where English is the primary language, neglecting to consider the cultural and linguistic nuances of other societies. In this paper, we aim to explore the inherent biases and toxicity of LLMs, specifically within the social context of Korea. We devise a set of prompts that reflect major societal issues in Korea and assign varied personas to both ChatGPT and GPT-4 to assess the toxicity of the generated sentences. Our findings indicate that certain personas or prompt combinations consistently yield harmful content, highlighting the potential risks associated with specific persona-issue alignments within the Korean cultural framework. Furthermore, we discover that GPT-4 can produce more than twice the level of toxic content than ChatGPT under certain conditions.

Counter-narrative generation, i.e., the generation of fact-based responses to hate speech with the aim of correcting discriminatory beliefs, has been demonstrated to be an effective method to combat hate speech. However, its effectiveness is limited by the resource-intensive nature of dataset construction processes and only focuses on the primary language. To alleviate this problem, we propose a Korean Hate Speech Counter Punch (KHSCP), a cost-effective counter-narrative generation method in the Korean language. To this end, we release the first counter-narrative generation dataset in Korean and pose two research questions. Under the questions, we propose an effective augmentation method and investigate the reasonability of a large language model to overcome data scarcity in low-resource environments by leveraging existing resources. In this regard, we conduct several experiments to verify the effectiveness of the proposed method. Our results reveal that applying pre-existing resources can improve the generation performance by a significant margin. Through deep analysis on these experiments, this work proposes the possibility of overcoming the challenges of generating counter-narratives in low-resource environments.

Translating major language resources to build minor language resources becomes a widely-used approach. Particularly in translating complex data points composed of multiple components, it is common to translate each component separately. However, we argue that this practice often overlooks the interrelation between components within the same data point. To address this limitation, we propose a novel MT pipeline that considers the intra-data relation. in implementing MT for training data. In our MT pipeline, all the components in a data point are concatenated to form a single translation sequence and subsequently reconstructed to the data components after translation. We introduce a Catalyst Statement (CS) to enhance the intra-data relation, and Indicator Token (IT) to assist the decomposition of a translated sequence into its respective data components. Through our approach, we have achieved a considerable improvement in translation quality itself, along with its effectiveness as training data. Compared with the conventional approach that translates each data component separately, our method yields better training data that enhances the performance of the trained model by 2.690 points for the web page ranking (WPR) task, and 0.845 for the question generation (QG) task in the XGLUE benchmark.

Byte Pair Encoding is an effective approach in machine translation across several languages. However, our analysis indicates that BPE is prone to over-segmentation in the morphologically rich language, Korean, which can erode word semantics and lead to semantic confusion during training. This semantic confusion, stemming from over-segmentation, ultimately contributes to a degradation of overall translation quality. To address this issue, we introduce Length-aware Subword Vocabulary Construction (LeVoC), a novel approach strategically incorporating longer words into the vocabulary. By utilizing an external monolingual Korean corpus, LeVoC extracts and integrates long words, effectively preserving morphological information and reducing semantic confusion. Our experiments demonstrate that LeVoC not only significantly outperforms BPE, but also can be applied to and surpass current state-of-the-art morpheme-aware subword tokenization methods. We provide evidence that the difficulty in translating sentences with long words in Korean is associated with morphological compositionality, and LeVoC’s ability to reduce semantic confusion during training leads to improved translation quality.

Rather than continuing the conversation based on personalized or implicit information, the existing conversation system generates dialogue by focusing only on the superficial content. To solve this problem, FoCus was recently released. FoCus is a persona-knowledge grounded dialogue generation dataset that leverages Wikipedia’s knowledge and personal persona, focusing on the landmarks provided by Google, enabling user-centered conversation. However, a closer empirical study is needed since research in the field is still in its early stages. Therefore, we fling two research questions about FoCus. “Is the FoCus whether for conversation or question answering?” to identify the structural problems of the dataset. “Does the FoCus model do real knowledge blending?” to closely demonstrate that the model acquires actual knowledge. As a result of the experiment, we present that the FoCus model could not correctly blend the knowledge according to the input dialogue and that the dataset design is unsuitable for the multi-turn conversation.