General General News

AI Models Face Challenges in Balancing Complex and Simple Tasks

Seoul:Large language models have demonstrated remarkable proficiency in solving complex mathematical problems but continue to struggle with simpler tasks, such as reading an analogue clock. This discrepancy highlights ongoing challenges in artificial intelligence (AI).

According to World Economic Forum, Google's DeepMind's Gemini Deep Think achieved a gold medal at the 2025 International Mathematical Olympiad by scoring 35 points, showcasing its ability to think in natural language. However, the same generation's best model for visual reasoning, tested on ClockBench, correctly read clocks only 50.6% of the time, compared to 90.1% for humans.

This inconsistency underscores the concept of 'jagged intelligence,' where AI models excel in certain tasks but fail in others. Such unpredictability creates operational risks for businesses relying on AI in their workflows, as models can introduce errors or misunderstandings.

Neurosymbolic AI, combining neural networks and symbolic AI, is emerging as a solution to enhance AI reliability. This approach allows for pattern recognition and logical reasoning, potentially improving AI's dependability in various applications.

The European Commission's AI regulations, in effect since 2 August 2026, emphasize the need for transparent AI systems. By integrating symbolic reasoning, AI can offer clearer decision-making processes, aligning with regulatory requirements.

As companies evaluate AI technologies, they must consider how models perform on relevant benchmarks and their ability to self-correct and explain their reasoning. This approach will be crucial as AI continues to evolve and expand its role in business operations.