AI framework cuts computing use by two-thirds while more than tripling success in virtual tasks

edited by Gaby Clark, reviewed by Robert Egan

This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:

Add to Preferred Sources OGBench. Credit: arXiv (2026). DOI: 10.48550/arxiv.2610.07899 A research team led by professor Minhae Kwon of the School of Electronic and Electrical Engineering has developed two core next-generation artificial intelligence (AI) technologies that enable AI systems to handle complex, long-horizon tasks more effectively while reducing computational costs and improving safety. The findings address key challenges that have hindered the commercialization of costly AI assistant services as well as the deployment of AI in safety-critical areas such as robotics and autonomous driving. The findings are published on the arXiv preprint server.

AI assistants that receive user instructions and autonomously plan and execute tasks, such as ChatGPT-based agents, are becoming increasingly widespread. However, many existing AI agents rely on repeated feedback from another high-performance AI model at every stage of a task—essentially asking, "Am I doing this correctly?" at each step. This process significantly increases server and computational costs while also slowing response times.

To address this issue, the research team, with Ph.D. student Heewon Park as first author, developed a selective-critique framework called SAG (Self-improving Agent with Gated critique). Much like a student who studies independently and asks a teacher for help only when encountering a particularly difficult problem, SAG allows an AI agent to assess its own uncertainty and seek external feedback only when it is likely to be worth the cost. The framework also enables the AI to accumulate successful experiences and progressively improve on its own, reducing its dependence on external feedback over time.

As a result, the system reduced computational resource consumption to approximately one-third that of conventional approaches while increasing the task success rate in complex virtual environments from 24.6% to 78.3%, representing more than a threefold improvement.

In parallel, the research team, with master's student Guhyeon Kang as first author, developed VAN-Flow, a high-reliability reinforcement learning technology designed to prevent AI systems from mistaking chance successes for genuinely effective strategies in safety-critical environments such as robotics and autonomous vehicles.

When AI is trained solely on previously collected data because real-world trial and error could lead to accidents, the system may occasionally learn from risky actions that happened to succeed by chance. This can cause the AI to incorrectly regard unstable behavior as an effective strategy, potentially resulting in system failures or malfunctions.

The team devised a new mathematical algorithm that filters out unreliable behaviors dependent on chance and instead prioritizes actions that produce consistently successful outcomes across varying conditions. Based on this approach, VAN-Flow achieved higher success rates than most existing state-of-the-art methods across more than 40 complex robotic-control simulation tasks involving significant environmental disturbances and long-horizon precision control.

Both studies by the SKKU research team were accepted for the 40th Conference on Neural Information Processing Systems ( NeurIPS 2026 ), where leading researchers and global technology companies such as Google and OpenAI gather to present advances in the field.

Kwon said, "It is particularly meaningful that our student researchers took the lead in developing practical foundational technologies that can be readily applied to real-world and industrial environments."

She added, "We will continue contributing to the development of next-generation autonomous AI systems that businesses and individuals can rely on with confidence, without excessive computational costs."

Heewon Park et al., Selective Critique for Cost-Aware LLM Agents in Long-Horizon Decision Making, arXiv (2026). DOI: 10.48550/arxiv.2610.07335 Guhyeon Kang et al., Variance-Averse n-Step Offline Reinforcement Learning for Sparse Long-Horizon Environments, arXiv (2026). DOI: 10.48550/arxiv.2610.07899

MA in English, copy editor since 2021 with experience in higher education and health content. Dedicated to trustworthy science news. Full profile →

Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile →