We conduct full-cycle research on language models—covering pre-training, post-training, and fine-tuning—focused on enabling machines to understand and generate human language.
Our goal is to develop game-specialized language model agents capable of comprehension, reasoning, decision-making, and emotional expression. We explore a wide spectrum of models, from large-scale LLMs to small, lightweight sLMs suitable for on-device deployment.
KRAFTON AI seeks to create natural and emotionally resonant human-computer interactions that enrich player immersion.

Our multi-modal learning research focuses on enabling AI to understand and process diverse data types—such as images, text, speech, and video—in an integrated manner. We aim to fully leverage the unique information each modality provides, for example, by combining images and text to build thematic databases, or merging language models with TTS systems to generate contextually appropriate speech. Ultimately, we envision multi-modal AI that can comprehend real-time video, voice, and text during gameplay, communicate with players via voice, and play games alongside them.
At KRAFTON AI, we are committed to going beyond siloed AI fields—fusing technologies to solve complex, ambitious challenges and realizing immersive human-AI interaction through seamless integration of all AI modalities.

Reinforcement learning is a technique where an agent is developed to interact with its environment so it can learn how to find the optimal action strategy (policy). This technology is closest to KRAFTON’s core drive as a game developer and maker of masterpieces.
For example, we could train the agent on important rules and strategies from a game and have it compete with other users or help developers test the game as it is being designed.
Reinforcement learning will play a crucial role in interactions between humans and computers as it helps AI develop the human-like capacity to learn and make decisions about complicated problems.

At KRAFTON AI, we research generative AI technologies to automate and enhance the creation of visual and audio assets essential for game development. Our work includes inferring and modeling 3D structures from 2D images, reconstructing character motion from human video footage (Video-to-Motion), and generating emotionally expressive speech from text (Text-to-Speech).
We also explore Voice Conversion technologies that can naturally transform one person’s voice into another. Through these innovations, we aim to maximize both production efficiency and immersive gameplay experiences.

Data-centric approaches focus on the quality of data used to enhance model performance. We study data management strategies surrounding data collection, processing, labeling, augmentation, and other processes that play an important role in building high-quality datasets.
For example, we study whether it is possible to achieve the same level of performance with less data or what type of data we would need to supplement for improved model performance.
Data-centric research based on deep learning enables AI models to make more accurate and powerful predictions. This field contributes to improving the reliability and fairness of AIs and is crucial in establishing a long-term direction and foundation for research organizations.