KISTI’s KONI Team Has Two Papers Accepted at ICLR, Securing Core Technologies for Sovereign AI Foundation Model Updates

- Achieves state-of-the-art Korean reasoning performance and introduces hallucination mitigation techniques

- Advances AI innovation for science through the development of an AI Co-Scientist

□ The Korea Institute of Science and Technology Information (KISTI, President Sik Lee) announced that two research papers by its KONI (KISTI Open Neural Intelligence) research team—developing a science- and technology-specialized large language model (LLM)—have been accepted at the International Conference on Learning Representations (ICLR), one of the world’s most prestigious conferences in artificial intelligence. The achievement underscores the growing global competitiveness of Korean-language LLMs.

□ Alongside NeurIPS (Conference on Neural Information Processing Systems) and ICML (International Conference on Machine Learning), ICLR is widely regarded as one of the “Big Three” AI conferences, recognized for cutting-edge research in deep learning and representation learning. It is also a major venue closely followed by global technology leaders such as Google, Meta, and OpenAI.

□ In collaboration with the non-profit open-source research group HAERAE, led by Gyujin Son, the KONI team carried out a project to develop a Korean-focused reasoning model. As part of this effort, the team constructed the Yi-SANG training dataset, comprising 5.79 million native Korean prompts and 3.7 million long-form reasoning trajectories. To date, Yi-SANG represents the largest publicly available post-training dataset for Korean language models.

□ The KONI team also introduced a Language-Mixed Chain-of-Thought (CoT) approach, in which reasoning is conducted in English while final responses are generated in Korean. This strategy overcomes the logical limitations of Korean-only models, minimizes translation artifacts, and significantly improves reasoning efficiency. Models trained with this approach achieved state-of-the-art Korean reasoning performance, surpassing global models of comparable scale, including DeepSeek-R1-32B.

□ In addition, the team developed LoRA-Gated Contrastive Decoding (LGCD), a novel technique designed to address catastrophic forgetting, a common challenge in adapting models to specific languages or domains. LGCD operates solely at inference time without requiring additional model training, dynamically extracting and correcting internal knowledge to improve factual consistency. The technique is particularly effective in suppressing hallucinations in specialized domains where high accuracy is critical.

□ These achievements provide a key technological foundation for updating KONI based on the Sovereign AI Foundation Model, in line with the policy direction recently emphasized by Kyunghoon Bae, Deputy Prime Minister and Minister of Science and ICT during the Ministry’s policy briefing.

□ Building on this research, the KONI team plans to advance the development of an AI Co-Scientist—an intelligent research agent designed to support scientists by formulating hypotheses, analyzing experimental data, and collaborating on complex scientific problems. Through this effort, KISTI aims to strengthen national AI sovereignty and contribute to the transformation of Korea’s research ecosystem toward AI for Science.

□ President Sik Lee of KISTI stated, “This achievement demonstrates the global competitiveness of Korean-language AI technologies. We will continue to advance KONI to lead innovation in AI for Science and the development of AI Co-Scientists.”