A team led by Professor Minha Kwon of the School of Electronic and Electrical Engineering at Sungkyunkwan University (Republic of Korea) has presented two technologies for artificial intelligence (AI) systems. They are meant to help AI handle long and complex tasks, use fewer computing resources and operate more safely. The results were published on the arXiv preprint server, a platform where scientific papers are posted before journal peer review.
One problem with modern AI assistants that plan and carry out a user's instructions on their own is constant double-checking. Many such systems at every step ask another, more powerful model: "Am I doing everything right?" This makes server computing more expensive and slows down the response.
The first development, SAG (Self-improving Agent with Gated critique), proposes that the AI assess its own uncertainty and request an external hint only when it is worth it. This is similar to a student who studies independently and goes to the teacher only with a difficult problem. The system also accumulates successful experience and over time depends less and less on outside hints.
In complex virtual environments this approach cut computing resource use to roughly one third of the usual level. The share of successfully completed tasks rose from 24.6% to 78.3%, more than threefold. Graduate student Hiwon Pak is listed as the first author of this work.
The second technology, VAN-Flow, was created for reinforcement learning, a method in which AI learns by trial and error, receiving conditional "rewards" for successful actions. Its first author is master's student KuHyun Kang. The development is intended for robotics and self-driving cars, where real-world experiments can lead to accidents, so AI is often trained only on pre-collected data.
Under such conditions the system may mistake random luck for a reliable strategy and start repeating risky behavior. The new mathematical algorithm filters out actions whose success depended on chance and gives priority to those that consistently produce results under different conditions. In more than 40 complex robot control simulations with strong interference and long precise operations, VAN-Flow showed a success rate higher than most existing state-of-the-art methods.
Both works were accepted to the 40th NeurIPS 2026 conference, which brings together researchers and representatives of technology companies, including Google and OpenAI. "It is especially significant that our student researchers led the development of practical foundational technologies that can be quickly applied in real and industrial environments," the press service quoted Kwon as saying. She added that the team will continue working on autonomous AI systems that businesses and people can trust without excessive computing costs.