Selectstar said on Sept. 28 that 5 papers involving its researchers on AI reliability and safety were accepted at the natural language processing (NLP) conference EMNLP 2026.

The accepted papers include 1 for the main conference, 1 for the official ancillary proceedings (Findings) and 3 workshop papers. EMNLP 2026 will be held in Budapest, Hungary, from Oct. 24 to 29.

The study accepted for the main conference presented an evaluation method that analyses not only whether AI gets questions right but also whether, when it is wrong, it chooses incorrect answers in a way similar to humans. It compares the distribution of correct and incorrect answer choices by humans and large language models (LLMs) to measure how closely AI resembles human judgment patterns.

The paper accepted in the proceedings analysed the decision reliability of guard models that filter harmful requests. It confirmed that some guard models can make wrong decisions while showing high confidence even when judgments are uncertain. Selectstar plans to use this in guardrail technology aimed at reducing excessive blocking in AI services.

For workshops, 2 studies in finance and 1 study on AI safety by cultural sphere were accepted. The finance research covered a way to improve AI guardrails while maintaining existing defense performance, and a technology to detect financial phishing sites by analysing multiple webpages together.

Selectstar plans to apply the results to its AI evaluation platform Datumo as well as red teaming and guardrail technologies to advance AI safety evaluation technology by country and industry.

Selectstar CEO Se-yup Kim (김세엽) said, "We will link research outcomes to evaluation, red teaming and guardrail technologies and develop them into an AI reliability verification system that companies can use in real service environments."