Large language models leverage both parametric knowledge acquired during pretraining and in-context knowledge provided at inference time. Crucially, when these sources conflict, models arbitrate based on their internal confidence, preferring parametric knowledge for high-confidence facts while deferring to context for less familiar ones. However, the training conditions that give rise to these fundamental behaviors remain unclear. Here we conduct controlled experiments using synthetic corpora to identify the specific data properties that shape knowledge utilization. Our results reveal a counterintuitive finding: the robust, balanced use of both knowledge sources is an emergent property that requires the co-occurrence of three factors typically considered detrimental, including (i) intra-document repetition, (ii) a moderate degree of intra-document inconsistency, and (iii) a skewed knowledge distribution. We further show that these dynamics arise in real-world language model pretraining and analyze how post-training procedures reshape arbitration strategies. Together, our findings provide empirical guidance for designing training data that supports the reliable integration of parametric and in-context knowledge in language models.

Open-weight large language models (LLMs) enable broad customization, but also increase exposure to post-release misuse, including malicious fine-tuning (MFT). To mitigate this risk, many prior defenses aim to improve the robustness of open-weight models to MFT by constraining adversarial fine-tuning dynamics in parameter space or mitigating harmful information encoded in internal representations. Nevertheless, since malicious fine-tuning can still erode safety, developing robust safeguards for open-weight models that fundamentally mitigate this risk remains an open research problem. In this paper, we characterize a safety region for open-weight LLMs and propose Safety Guidance Trigger (SGT), which guides fine-tuning toward the safety manifold to preserve alignment. SGT has two stages: (1) optimizing a safety trigger that steers the base model toward safe responses and (2) training the open-weight model to align its internal features with trigger-induced safety representations. We demonstrate that SGT substantially improves robustness against malicious fine-tuning, requiring adversaries to increase their data budget significantly to compromise safety. Our analysis shows that SGT anchors model representations to a safety region, which remains stable under malicious fine-tuning.

Human moral judgment is context-dependent and changes based on interpersonal relationships. As large language models (LLMs) increasingly serve as decision-support systems, it is critical to understand if they encode these social nuances. We characterize LLM behavior using the Whistleblower’s Dilemma, systematically varying two experimental factors: crime severity and relational closeness. Our study compares three evaluative perspectives: (1) moral rightness (general prescriptive norms), (2) predictive human behavior (how models expect people to navigate social situations), and (3) models’ own decision-making. By analyzing the reasoning processes, we find a clear cross-perspective divergence: moral rightness remains consistently fairness-oriented, while predicted human behavior shifts with relational context toward loyalty. Crucially, the model decisions mirror moral rightness judgments, rather than their behavioral predictions. This cross-perspective inconsistency suggests that LLM decision-making favors abstract rules over the social sensitivity found in their internal modeling, potentially producing conflicting expectations in real-world deployments.

Humans have an inherent need for community belongingness. This paper investigates this fundamental social motivation by compiling a large collection of parallel datasets comprising over 7 million posts and comments from Reddit and 200,000 posts and comments from Dread, a dark web discussion forum, covering similar topics. Grounded in five theoretical aspects of the Sense of Community framework, our analysis indicates that users on Dread exhibit a stronger sense of community membership. Our data analysis reveals striking similarities in post content across both platforms, despite the dark web’s restricted accessibility. However, these communities differ significantly in community-level closeness, including member interactions and greeting patterns that influence user retention and dynamics. We publicly release the parallel community datasets for other researchers to examine key differences and explore potential directions for further study.

While detecting offensive language in online spaces remains an important societal issue, there is still a significant gap in existing research and practial datasets specific to chatbots. Furthermore, many of the current efforts by service providers to automatically filter offensive language are vulnerable to users’ deliberate text manipulation tactics, such as misspelling words. In this study, we analyze offensive language patterns in real logs of 6,254,261 chat utterance pairs from the commercial chat service Simsimi, which cover a variety of conversation topics. Based on the observed patterns, we introduce a novel offensive language detection method—a contrastive learning model that embeds chat content with a random masking strategy. We show that this model outperforms existing models in detecting offensive language in open-domain chat conversations while also demonstrating robustness against users’ deliberate text manipulation tactics when using offensive language. We release our curated chatbot dataset to foster research on offensive language detection in open-domain conversations and share lessons learned from mitigating offensive language on a live platform.

Understanding the interplay between emotions in language and user behaviors is critical. We study how moral emotions shape the political participation of users based on cross-cultural online petition data. To quantify moral emotions, we employ a context-aware NLP model that is designed to capture the subtle nuances of emotions across cultures. For model training, we construct and share a moral emotion dataset comprising nearly 50,000 petition sentences in Korean and English each, along with emotion labels annotated by a fine-tuned LLM. We examine two distinct types of user participation: general support (i.e., registered signatures of petitions) and active support (i.e., sharing petitions on social media). We discover that moral emotions like other-suffering increase both forms of participation and help petitions go viral, while self-conscious have the opposite effect. The most prominent moral emotion, other-condemning, led to polarizing responses among the audience. In contrast, other-praising was perceived differently by culture; it led to a rise in active support in Korea but a decline in the UK. Our findings suggest that both moral emotions embedded in language and cultural perceptions are critical to shaping the public’s political discourse.

Quotes are critical for establishing credibility in news articles. A direct quote enclosed in quotation marks has a strong visual appeal and is a sign of a reliable citation. Unfortunately, this journalistic practice is not strictly followed, and a quote in the headline is often “contextomized.” Such a quote uses words out of context in a way that alters the speaker’s intention so that there is no semantically matching quote in the body text. We present QuoteCSE, a contrastive learning framework that represents the embedding of news quotes based on domain-driven positive and negative samples to identify such an editorial strategy. The dataset and code are available at https://github.com/ssu-humane/contextomized-quote-contrastive.

The potential social harms that large language models pose, such as generating offensive content and reinforcing biases, are steeply rising. Existing works focus on coping with this concern while interacting with ill-intentioned users, such as those who explicitly make hate speech or elicit harmful responses. However, discussions on sensitive issues can become toxic even if the users are well-intentioned. For safer models in such scenarios, we present the Sensitive Questions and Acceptable Response (SQuARe) dataset, a large-scale Korean dataset of 49k sensitive questions with 42k acceptable and 46k non-acceptable responses. The dataset was constructed leveraging HyperCLOVA in a human-in-the-loop manner based on real news headlines. Experiments show that acceptable response generation significantly improves for HyperCLOVA and GPT-3, demonstrating the efficacy of this dataset.

Two types of topic modeling predominate: generative methods that employ probabilistic latent models and clustering methods that identify semantically coherent groups. This paper newly presents UTopic (Unified neural Topic model via contrastive learning and term weighting) that combines the advantages of these two types. UTopic uses contrastive learning and term weighting to learn knowledge from a pretrained language model and discover influential terms from semantically coherent clusters. Experiments show that the generated topics have a high-quality topic-word distribution in terms of topic coherence, outperforming existing baselines across multiple topic coherence measures. We demonstrate how our model can be used as an add-on to existing topic models and improve their performance.

This paper presents a time-topic cohesive model describing the communication patterns on the coronavirus pandemic from three Asian countries. The strength of our model is two-fold. First, it detects contextualized events based on topical and temporal information via contrastive learning. Second, it can be applied to multiple languages, enabling a comparison of risk communication across cultures. We present a case study and discuss future implications of the proposed model.

This study analyzes the political slants of user comments on Korean partisan media. We built a BERT-based classifier to detect political leaning of short comments via the use of semi-unsupervised deep learning methods that produced an F1 score of 0.83. As a result of classifying 21.6K comments, we found the high presence of conservative bias on both conservative and liberal news outlets. Moreover, this study discloses an asymmetry across the partisan spectrum in that more liberals (48.0%) than conservatives (23.6%) comment not only on news stories resonating with their political perspectives but also on those challenging their viewpoints. These findings advance the current understanding of online echo chambers.