Recent large reasoning models produce explicit reasoning traces before final answers, yet their behavior in culturally grounded reasoning remains insufficiently understood. In this work, we identify cultural overriding, a step-level failure mode in which a model initially recognizes a relevant local cultural norm but later weakens or replaces it with a more general normative frame, resulting in an incorrect judgment. To address this failure, we introduce CulturePRM, a 1B-parameter process reward model trained on step-annotated synthetic cultural reasoning traces to verify whether intermediate steps preserve, override, or are unrelated to the target cultural norm. At inference time, CulturePRM aggregates step-level rewards to rerank candidate reasoning traces via Best-of-N selection.

We evaluated CulturePRM across multiple cultural reasoning benchmarks and base reasoning models and observed consistent improvements, with gains of up to 12.06 pp. These gains were not uniform across classes and were concentrated in lower-recall classes, while performance in already strong classes was largely preserved, thereby reducing class-wise disparities. A finer-grained error analysis further revealed substantial recovery among baseline errors alongside a marked reduction in cultural overriding. These findings suggest that culturally aligned reasoning requires models not only to recognize local cultural norms, but also to preserve them throughout the reasoning process.

We present a novel end-to-end personality-based synthetic dialogue data generation pipeline, specifically designed to elicit responses from large language models via prompting. We design the prompts to generate more human-like dialogues considering real-world scenarios when users engage with chatbots. We introduce PSYDIAL, the first Korean dialogue dataset focused on personality-based dialogues, curated using our proposed pipeline. Notably, we focus on the Extraversion dimension of the Big Five personality model in our research. Experimental results indicate that while pre-trained models and those fine-tuned with a chit-chat dataset struggle to generate responses reflecting personality, models trained with PSYDIAL show significant improvements. The versatility of our pipeline extends beyond dialogue tasks, offering potential for other non-dialogue related applications. This research opens doors for more nuanced, personality-driven conversational AI in Korean and potentially other languages.

Unsupervised text style transfer is a challenging task that aims to alter the stylistic attributes of a given text without affecting its original content. One of the methods to achieve this is controllable style transfer, which allows for the control of the degree of style transfer. However, an issue encountered with controllable style transfer is the instability of transferred text fluency when the degree of the style transfer changes. To address this problem, we propose a novel approach that incorporates additional syntax parsing information during style transfer. By leveraging the syntactic information, our model is guided to generate natural sentences that effectively reflect the desired style while maintaining fluency. Experimental results show that our method achieves robust performance and improved fluency compared to previous controllable style transfer methods.