OpenAI夺金IOI，但输给3位中国高中生

Core Viewpoint - OpenAI's reasoning model achieved a record score of 533.29 in the IOI competition, ranking sixth among 330 human participants and first among all AI competitors, showcasing significant advancements in AI capabilities [1][4][13]. Group 1: Performance Highlights - OpenAI's AI reasoning system surpassed 98% of participants, demonstrating a marked improvement compared to the previous year's model [13]. - The AI system did not utilize a new training model specifically for IOI but integrated multiple general reasoning models for the competition [2][11]. - The gold medal score for IOI was set at 438.30, with only 28 participants achieving this score out of 330 competitors from 84 countries [9]. Group 2: Competition Structure - The IOI competition requires participants to solve three high-difficulty algorithm problems independently over two days, with a strict 5-hour time limit each day and no internet access [8][10]. - OpenAI's AI system operated under the same rules as human participants, with a limit of 50 submissions [10][11]. Group 3: Model Development - OpenAI's model for IOI 2024, named o1-ioi, was specifically fine-tuned for programming tasks, yet it only scored 213 points, ranking in the 49th percentile [14][15]. - The AI system generated 10,000 candidate solutions for each sub-task and utilized a complex test-time reasoning strategy to select the final submissions [17].