LG's First Public Vision Language Model, EXAONE 4.5 - LG AI Research BLOG

LG AI Research introduced EXAONE 4.5, its first open-weight Vision Language Model (VLM). Moving beyond the existing language-centric architecture, the model has evolved into a true Multimodal model that understands text and images simultaneously by integrating a self-developed Visual Encoder.

The core of the model is its Native Multimodal Pretraining approach. Instead of simply combining vision and language modules, the two types of information are trained together from the start, maximizing multimodal understanding capability. In addition, specialized data for STEM and Document Understanding domains was built to enhance practicality in industrial settings.

The following designs were applied for technical optimization and improved inference speed.

• GQA (Grouped Query Attention): Applied to the attention mechanism within the Visual Encoder to reduce computation and memory usage while securing inference speed.
• MTP (Multi-Token Prediction): By predicting multiple subsequent tokens simultaneously rather than just a single token, inference speed was improved by about 1.5x or more compared to before.

During training, high-quality data including Thinking Trajectories was used to strengthen logical reasoning capability. In the post-training stage, a self-developed AGAPO algorithm was introduced. This algorithm guides the model to autonomously avoid logical errors and improves training efficiency through asymmetric sampling that finds learning signals even within incorrect answers, along with stable advantage calculation.