AI Framework Cuts Computing Use by Two-Thirds, Triples Virtual Task Success

Researchers at Sungkyunkwan University, led by Minhae Kwon of the School of Electronic and Electrical Engineering, developed two next-generation artificial intelligence technologies designed to make autonomous AI assistants and safety-critical AI systems more effective, efficient and reliable. The work, led by doctoral student Heewon Park and master's student Guhyeon Kang, addresses two major barriers to commercial AI deployment: excessive computational costs and the risk that AI systems learn unstable strategies in robotics or autonomous driving environments. The first technology, SAG, or Self-improving Agent with Gated critique, is a selective-critique framework for large language model agents. Many current AI assistants depend on continuous validation from another high-performance model, repeatedly asking whether each step of a task is correct. That approach improves oversight but increases server costs and slows response times. SAG instead allows an AI agent to evaluate its own uncertainty and request external feedback only when the cost is justified. Over time, the system builds on successful experiences and relies less on outside critique. In complex virtual environments, SAG cut computational resource consumption to about one-third of conventional approaches while raising task success rates from 24.6% to 78.3%, more than tripling performance. The second technology, VAN-Flow, is a high-reliability reinforcement learning method for safety-critical settings such as robotics and autonomous vehicles. When AI systems are trained using previously collected data rather than real-world trial and error, they can sometimes learn from risky actions that succeeded only by chance. That can cause them to treat unstable behavior as effective strategy. VAN-Flow uses a mathematical filtering method to remove unreliable, luck-dependent behaviors and prioritize actions that consistently succeed under changing conditions. Across more than 40 complex robotic-control simulations involving environmental disturbances and long-horizon precision tasks, VAN-Flow outperformed most state-of-the-art methods. Both studies were accepted to NeurIPS 2026, one of the leading venues for AI research, and were published as arXiv preprints in 2026. The results are significant for AI assistants, robotics and autonomous systems because they target two practical needs: reducing the cost of agentic AI while improving reliability in environments where errors can have serious consequences. The team said the work shows that student-led research can produce foundational technologies ready for real-world and industrial use.