OneLine AI: Multilingual AI Math Reasoning Accepted to ACL 2025 Main

OneLine AI, together with KAIST AI researchers, has opened a new horizon for global AI research with their research paper, 'Linguistic Generalizability of Test-Time Scaling in Mathematical Reasoning,' which significantly enhanced the multilingual mathematical reasoning capabilities of Large Language Models (LLMs). The paper has been officially adopted by the main conference of ACL 2025, the most prestigious academic conference in the field of AI. The core of this research is a deep analysis of whether 'Test-Time Scaling (TTS),' a technique that maximizes performance by investing additional computational resources when an AI model solves actual problems, maintains consistent reasoning capabilities across various language environments. This signifies a major advancement in terms of linguistic universality and efficiency for AI language models. Moving beyond the limitations of existing simple math problem benchmarks, the research team developed a new benchmark, 'MCLM (Multilingual Competition Level Math),' consisting of high-difficulty, competition-level math problems translated into 55 languages. This benchmark is designed to precisely evaluate an AI's ability to process complex mathematical reasoning in multiple languages. The experiments compared and analyzed three major TTS methods: Outcome Reward Modeling (ORM), Process Reward Modeling (PRM), and Budget Forcing (BF). Notably, OneLine AI's 'MR1-1.5B' model demonstrated outstanding performance with the Budget Forcing method. Despite being a lightweight model with only 1.5B parameters, it scored 30.93 points on the MCLM benchmark, significantly outperforming competing models of similar size, such as Qwen2.5-Math-1.5B-Instruct (23.98 points) and DeepSeek-R1-1.5B (28.83 points), thereby proving the potential of the TTS technique. Furthermore, multilingual performance verification confirmed that the TTS method is effective in language environments other than English. OneLine AI plans to contribute to the activation of subsequent research within the global AI community by releasing the MCLM benchmark dataset used in this study on the open-source platform HuggingFace. With the adoption by ACL 2025, OneLine AI is also accelerating the commercialization of its multilingual AI technology. The multilingual mathematical reasoning and TTS technologies accumulated through this research will be applied to its generative AI-based global investment information platform, 'Finola'. Finola is a financial assistant service that provides U.S. stock market information in 10 languages, supporting investors worldwide to gain sophisticated investment insights without language barriers, and is expected to present a successful model where AI research achievements translate into real market value.