Correct Answers for Some High-Difficulty Subjective Questions... Weaknesses in Geometry and Inference-Type Problems
On the 7th (U.S. local time), OpenAI officially announced its latest language model GPT-5. Image=OpenAI
The next-generation large language model (LLM) 'GPT-5', released by OpenAI on the 7th (local time), achieved scores on the College Scholastic Ability Test (CSAT) that would make it competitive for admission to top universities in the Seoul metropolitan area.
On the 8th, when the 2025 CSAT Korean (Speech and Composition), Mathematics (Calculus), and English questions were entered into the GPT-5 chatbot from number 1 to the last question and graded, it scored 95 in Korean, 82 in Mathematics, and 92 in English based on raw scores. Compared to last year's CSAT grade cutoffs, these scores correspond to the first grade for Korean and English, and the second grade for Mathematics.
The most notable change in this test was its mathematical calculation ability. GPT-5 was presented with mathematics questions 1 through 30 as images and produced answers to most of them within 1 to 2 seconds. The solution process was displayed using LaTeX formulas for improved readability. In particular, for one of the high-difficulty, 4-point subjective questions (numbers 29 and 30), it provided the correct answer and solution after about 1 minute and 30 seconds of calculation. However, it showed limitations with inference questions requiring complex condition comparisons or geometry questions requiring figure recognition, sometimes producing plausible but incorrect answers or responding that there was no answer.
In the Korean section, it answered all non-literary and composition questions correctly but made mistakes on the modern literature comparison (question 22) and classical poetry interpretation (question 34). In English, despite being given a chance to retry, it selected the same incorrect answers for the sentence order questions (numbers 37 and 43), resulting in a total of four incorrect answers.
In the inquiry section, GPT-5 performed strongly in social studies subjects. It solved even difficult questions accurately in subjects like Social and Cultural Studies and Ethics and Thought, but in science inquiry subjects such as Physics and Chemistry, which require chart analysis, it had a high error rate even on relatively easy questions.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
