Knowledge sharing • 2 min read •
The Illusion of Thinking: Evaluating Large Reasoning Models Through Problem Complexity
The "thinking AI" may not be thinking quite the way we believe. Apple's latest research suggests that state-of-the-art Large Reasoning Models (LRMs) break down once problems become sufficiently complex, while their apparent reasoning process scales in unexpected ways. The findings immediately sparked debate, with other researchers arguing that the real issue lies not in the models themselves, but in the experimental design. Are we seeing genuine intelligence—or simply the illusion of it?
With the latest generation of frontier language models, Large Reasoning Models (LRMs) have also emerged. Unlike conventional LLMs, these models generate detailed reasoning traces before producing an answer. While they consistently outperform standard language models on reasoning benchmarks, their fundamental capabilities, scaling properties, and limitations remain insufficiently understood. Current evaluation methods primarily rely on established mathematics and programming benchmarks, focusing almost exclusively on the accuracy of the final answer. However, this evaluation paradigm is often affected by data contamination and provides little insight into the structure and quality of the reasoning process itself.
Apple's widely discussed paper (Shojaee et al., 2025) systematically addresses these shortcomings by introducing controllable puzzle environments that allow researchers to precisely manipulate compositional complexity while maintaining consistent logical structures. This setup makes it possible not only to evaluate final answers but also to examine the internal reasoning traces of the models, providing a rare glimpse into how LRMs "think." Through extensive experiments across several puzzle domains, the researchers found that state-of-the-art LRMs experience complete accuracy collapse once task complexity exceeds a certain threshold. They also observed an unexpected scaling behavior: reasoning effort initially increased with problem complexity but later declined, despite the models still having sufficient token budgets available.
When comparing LRMs with standard LLMs under equivalent inference compute, the researchers identified three distinct performance regimes: (1) low-complexity tasks, where standard language models surprisingly outperform LRMs; (2) medium-complexity tasks, where additional reasoning provides a measurable advantage for LRMs; and (3) high-complexity tasks, where both model families experience complete performance collapse. The study concludes that LRMs exhibit fundamental limitations in precise computation: they fail to reliably execute explicit algorithms and display inconsistent reasoning behavior across different puzzle types.
Following the publication of Apple's paper, researchers from Anthropic and Open Philanthropy published a response raising substantial concerns about the experimental methodology (Opus & Lawsen, 2025). Their commentary, titled The Illusion of the Illusion of Thinking, argues that much of the reported reasoning collapse stems from flaws in the experimental design rather than inherent reasoning limitations. According to the authors, in the Towers of Hanoi experiments the models actually recognize token-length constraints and intentionally shorten their outputs instead of failing to reason correctly. They also point out that, under certain conditions, one of the river-crossing puzzles is mathematically unsolvable, yet the models were nevertheless evaluated as incorrect for failing to solve impossible instances. Furthermore, when the researchers changed the output representation—for example, asking the models to generate a Lua function instead of an explicit sequence of moves—the models achieved high accuracy even on Towers of Hanoi problems that had previously been classified as evidence of "collapse." This suggests that the observed limitations may be largely format-dependent rather than reflecting fundamental deficiencies in reasoning.
If you'd like to read the original study: Illusion of Thinking,
If you're also interested in the counterarguments: Commentary.
