One Line AI (CEO Jeong Han-eol) has developed ‘OLAF v2’, a Korean artificial intelligence (AI) model based on thought process and ‘test-time scaling’. Published on Hugging FaceIt was announced on the twenty first that it was done.
The thought process is a brand new reasoning method for AI models. This works in the same method to the present ‘Chain-of-Thought’. The model generates a pondering process that permits it to grasp a given problem and propose an answer. This makes it possible to systematically solve problems that require complex reasoning. On this process, errors are discovered and corrected on their very own, leading to a more reliable answer.
Test time scaling is a technique of allowing models to make use of more computational resources to take into consideration complex problems on their very own and perform deeper inferences. This helps the model reach the optimal solution by simulating various strategies and scenarios, and methods equivalent to ‘Monte Carlo Tree Search (MCTS)’ are typically used.
Olaf v2 was released in two versions: 14B and 1.5B. It has a specialized reasoning mode to unravel STEM problems equivalent to science, technology, and arithmetic.
It supports a context length of 32K and is designed for search augmented generation (RAG) and tool-based applications. Within the model learning process, we focused on repetitive data generation and a rejection mechanism for inappropriate questions to scale back hallucinations and increase reliability.

Because of this of the benchmark, it scored 91.96 points in GSM8K, a group of elementary and middle school math problems, and 36.20 points in Omni-MATH, a group of Olympiad-level problems.
This can be a higher rating than GPT-4o (GSM8K 91.21 points, Omni-Math 30.75 points). It is claimed that it is a meaningful result considering that the Olaf v2 model size is 14B.
CEO Jeong Han-eol said, “The outstanding reasoning ability demonstrated by Olaf v2 will show particular strengths in complex financial data evaluation and risk assessment,” adding, “We plan to develop it right into a specialized service for the financial sector through projects currently underway with major domestic financial firms.” said.
Reporter Park Soo-bin sbin08@aitimes.com