Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

Korean exam test: GPT and Gemini score near-perfect, domestic K-EXAONE slips on inquiry section

On South Korea's national college entrance exam, GPT and Gemini turned in near-perfect answers while the country's top domestic model, K-EXAONE, lost ground on the inquiry section. The comparison puts the gap between frontier models and a sovereign Korean model onto a single, publicly legible scorecard.

Published
韩国高考实测:GPT 与 Gemini 拿下高分,本土 K-EXAONE 在探究科目失分
Image source: 90daykorean.com

South Korea's national college entrance exam became a public capability test for AI this week. Local media reported that GPT and Gemini answered the exam at a near-perfect level, while K-EXAONE, the country's strongest domestic model, fell short on the inquiry section.

The comparison drew attention because the terms are unusually clean. The paper is fixed, the answers are unique and the marking scheme is public, so a global frontier model and a domestic one can be measured on the same questions instead of through subjective review.

Where K-EXAONE lost ground matters as much as the fact that it did. The inquiry section tests reading and reasoning over social-science and science material, which leans heavily on close reading of long Korean passages and on conventions specific to the local curriculum. That kind of capability is not reliably bought with English-language training data.

For the Korean AI industry, K-EXAONE has been positioned as a sovereign model, carrying expectations about local language, local industry and local regulation. Its standing on general leaderboards does not automatically carry over to hard, localized tasks.

Exam scores are a narrow probe: they reward structured problem solving under a fixed time limit, not the open-ended tool use and verification that real work demands. The testing conditions behind such results, including model versions, prompts and whether tools were allowed, are also rarely published in full.

What to watch next is whether K-EXAONE's developers treat the inquiry gap as a data problem or a reasoning problem, and whether Korean education and testing authorities reopen the question of how assessments should distinguish human work from machine work now that a chatbot can sit the same paper.

Why it matters

A domestic model's weaker localized reasoning is now publicly quantified, putting pressure on Korea's sovereign-model narrative to prove itself on Korean-language tasks rather than general leaderboards.

K-EXAONEBenchmarkKorea
Back to realtime news

Nearby Updates

All

11/20, 09:07

Huawei Cloud releases Ascend-native embodied world model CloudRobo-WM1.2

Huawei Cloud used an embodied-AI forum to unveil CloudRobo-WM1.2, described as an Ascend-native embodied world model, alongside its one-stop CloudRobo development platform. The platform reports more than a million navigation trajectories, over 50,000 manipulation trajectories and more than 30 third-party models, aimed at sparing robotics teams from rebuilding infrastructure from scratch.

11/20, 08:43

Anthropic's seven founders cashed out nearly $2B in record-breaking IPO

Anthropic's seven co-founders each took home about $280 million at the company's $83-per-share IPO price, for a combined payout of nearly $2 billion, TechCrunch reports. The offering was described as record-breaking, with the company raising close to $105 billion.

11/20, 10:34

Hainan releases its first filed tourism LLM as Yuduoduo agent lands in Rongshanliao

Hainan released its first filed large model for the culture and tourism sector, with an agent called Yuduoduo deployed in Rongshanliao on November 18, according to Guandian. The launch puts a locally filed model and agent into a concrete travel scenario rather than a generic chatbot.

11/20, 10:36

Kimi-K2.5 lands in Moonshot AI’s official Hugging Face repo as an image-text model

Moonshot AI has refreshed its official Hugging Face repository with Kimi-K2.5, a model whose pipeline is tagged image-text-to-text and whose weights ship in safetensors format. The repo reports 391,280 downloads and 2,902 likes, giving the new version a public multimodal footprint before any technical report appears.