Korean researchers have successfully integrated quantum computing, considered a next-generation game changer, into artificial intelligence (AI) training, creating a turning point that resolves the "explosion of computational costs" that has been the biggest obstacle for academia and industry. By boosting the training efficiency of "Robust AI"—which does not malfunction even under sudden environmental changes—by up to five times compared to existing methods, the commercialization of future industrial ecosystems tied to safety, such as self-driving cars and advanced robotics, is expected to accelerate further.
A research team led by Professor Yoon Sung-hwan of the Graduate School of Artificial Intelligence at the Ulsan National Institute of Science and Technology (UNIST) and a team led by Professor Kim Joong-heon of the Department of Electrical Engineering at Korea University announced Tuesday that they had developed "QRIM (Quantum Robust Inner Minimization)," a proprietary training technique that resolves a critical bottleneck in robust reinforcement learning based on quantum algorithms.
This research achievement was formally accepted at the "International Conference on Machine Learning (ICML) 2026," the most prestigious conference that draws the greatest attention from AI researchers worldwide, earning recognition for its world-class technological capability. In particular, among the quantum AI research accepted as main conference papers at ICML 2026, held this year at COEX in Seoul, this is the only achievement led by a Korean research team, which is assessed as raising the status of K-AI another notch.
Existing AI training systems, particularly "reinforcement learning"—which mimics human learning methods and shows excellent performance in sequential decision-making—carried a critical weakness. This is the phenomenon in which performance plummets when minor environmental changes or sudden variables not experienced during training occur. A representative example is when a self-driving car trained only on clear-weather road data encounters unexpected heavy snow or icy roads, driving along a completely wrong route or falling into an uncontrollable state.
To address this, academia has focused on "Robust Reinforcement Learning" research, which preemptively assumes the "worst-case scenarios" a system may face and repeatedly trains it to prepare for them. However, the more the precision of environmental variables was increased for safety, the more it was blocked by a "computational wall," where the number of cases a computer must calculate increases exponentially. Because all possible risk candidates must be verified one by one at every moment, physical training time became unrealistically long, making application in industrial settings nearly impossible.
The joint research team broke through this chronic computational bottleneck using "superposition," a core physical principle of quantum computing. Because quantum computers can hold the states of 0 and 1 simultaneously, it is possible to represent numerous environmental candidates within a single quantum state at once. Whereas traditional classical-computer-based AI lined up all scenarios in a single row and calculated them one by one to find the worst variable, the QRIM technique changed the paradigm by floating numerous future risk factors in quantum space all at once and scanning them at a glance.
The QRIM technique devised by the research team organically combines two core quantum algorithms within the reinforcement learning structure. First, it introduces the "Quantum Amplitude Estimation (QAE)" algorithm to estimate at ultra-high speed the expected reward value of the AI in each variable environment. Next, it activates the "Quantum Minimum Finding (QMF)" algorithm to capture the "worst environmental condition" most unfavorable to the AI among vast amounts of data.
The mathematical power of this technique is overwhelming. Breaking away from a structure in which a classical computer had to exhaustively examine N risk scenarios, it finds the exact same worst environment with only √N confirmations, at the square-root level. In an extreme situation where there are 10,000 sudden scenarios to review, a classical computer requires 10,000 operations, but with the QRIM module installed, the search is completed with just 100 operations.
Through discrete and continuous control simulation environments, the research team quantitatively verified that the QRIM technique achieves perfect robustness with only 20-30% of the computational load of existing classical methods. In particular, rather than remaining in a virtual simulator, they transplanted the algorithm onto IBM's 127-qubit quantum computer hardware used in actual industrial settings, ultimately proving its real operating performance. They confirmed for the first time in the world that the AI's malfunction-prevention function is stably maintained even in an environment where "hardware noise"—which inevitably occurs in actual quantum devices—is present.
Another strength of the QRIM technique lies in its overwhelming "versatility" and "portability." Lee Hyun-kyu, the first author of the paper, explained, "This is a technology newly designed with a structure that can accelerate the 'worst-case search' process—the biggest bottleneck in robust reinforcement learning—with quantum algorithms." He added, "Since existing reinforcement learning algorithms can be left as they are and only this search part replaced with a quantum module, it can be broadly applied to various reinforcement learning fields."
Professor Yoon Sung-hwan of UNIST expressed his expectations, saying, "This research is a concrete example showing how quantum computing can supplement the limitations of existing artificial intelligence," and "it could be utilized in fields such as robotics and self-driving, where adaptation to environmental changes is essential."
Meanwhile, this research was carried out with support from the Ministry of Science and ICT through the National Research Foundation (NRF)'s "Mid-Career Researcher Program," and the Institute of Information & Communications Technology Planning & Evaluation (IITP)'s "AI Graduate School Support Program," "AI Star Fellowship Program," "Regional Intelligence Innovation Talent Development Program," and "SW Star Lab" program.
