A new study co-authored by Apple researchers reveals that today’s most advanced large reasoning models (LRMs) struggle significantly when faced with increasingly complex tasks. While these AI models excel in solving moderate problems using step-by-step reasoning, their performance collapses entirely once task complexity crosses a certain threshold.
Unlike traditional models, LRMs are designed to mimic human-like thinking by generating internal reasoning steps before delivering answers. This approach has led to impressive gains in tasks requiring logical deduction. However, the new research shows these models hit a critical wall under higher cognitive loads.
To test the boundaries, researchers used controlled puzzle environments—such as the Tower of Hanoi, River Crossing, Blocks World, and Checkers Jumping—where they could gradually increase the difficulty while keeping rules constant. The findings were stark: as challenges intensified, LRM performance dropped sharply, often reaching zero correct outputs.
The study also observed a concerning pattern. As puzzles became harder, the models reduced their reasoning steps rather than expanding them—an unexpected sign of cognitive retreat. Interestingly, in simpler tasks, the models sometimes overthought the problem, missing the correct solution despite identifying it early in the process.
Researchers concluded that while step-by-step reasoning temporarily delays failure, it cannot override the underlying limitations of current AI systems. The report raises critical concerns about whether merely extending the reasoning process will be enough to create truly general AI capable of solving deeply complex, novel problems.
Explore More Cutting-Edge Tech Like This – Stay ahead of the game. Visit IT Tech News for more insights on AI, data tech, and innovation shaping our future.
News Source: Pymnts.com
