Introduction

Large reasoning models have been hailed as a breakthrough in artificial intelligence, promising to solve complex problems by thinking through them step by step before delivering an answer. Systems such as OpenAI's o-series, DeepSeek-R1, and Claude with thinking modes are designed to generate long internal chains of thought, self-reflect, and work through difficult logic. But a new study suggests these capabilities may have a hard ceiling. When puzzles become sufficiently complex, even the most advanced reasoning models experience a total collapse in accuracy, raising serious questions about whether current AI truly reasons or simply mimics the appearance of reasoning.

What Happened

Researchers tested frontier reasoning models against standard language models using four controllable puzzle environments: Tower of Hanoi, Checkers Jumping, River Crossing, and Blocks World. Unlike static math benchmarks, these puzzles allowed the team to precisely adjust complexity while keeping the underlying logic consistent. The results revealed three distinct performance regimes. At low complexity, standard non-reasoning models were often more accurate and token-efficient. At medium complexity, reasoning models pulled ahead. But at high complexity, both types of models collapsed to near-zero accuracy.

Perhaps more surprising was how the models behaved as they approached that breaking point. Instead of using more thinking tokens to tackle harder problems, reasoning models actually reduced their inference effort just as complexity peaked, despite having plenty of token budget remaining. Analysis of their internal thought traces showed additional weaknesses: models sometimes found correct answers early only to abandon them for incorrect alternatives, and at high complexity they frequently fixated on early mistakes and could not recover. Even when researchers handed the models explicit solving algorithms, performance did not meaningfully improve, suggesting a fundamental struggle with exact computation and logical execution.

Why This Matters

These findings challenge the assumption that simply giving AI models more time to think will reliably produce better results on hard problems. If reasoning models cannot follow prescribed algorithms or scale their thinking effort as difficulty increases, their real-world usefulness for complex planning, coding, and decision-making tasks may be more limited than benchmark scores suggest. The research also highlights flaws in how AI is evaluated. Popular math and coding benchmarks can suffer from data contamination and do not reveal how models reason internally. Controlled puzzle environments offer a clearer window into model behavior, but they also expose gaps that standard leaderboards hide.

For developers and businesses deploying AI, the implications are practical. A model that performs well on simple reasoning tasks may fail catastrophically once a problem crosses a complexity threshold, with no graceful degradation. Understanding where that threshold lies—and why models stop using available compute when they need it most—is essential for building reliable systems and setting realistic expectations about machine reasoning.

Key Takeaways

  • Advanced reasoning models experience complete accuracy collapse on puzzles beyond a certain complexity threshold.
  • Standard language models can outperform reasoning models on simple tasks, while reasoning models excel only at moderate complexity.
  • As problems approach the collapse point, reasoning models counterintuitively reduce their thinking effort despite having unused token capacity.
  • Internal thought traces show models often fixate on early errors and struggle to self-correct at high complexity.
  • Providing explicit algorithms does not prevent failure, indicating limitations in exact computation and logical execution.
  • Current benchmark-driven evaluations may overstate reasoning capabilities because they lack controlled complexity and trace-level analysis.

Conclusion

The idea that AI reasoning will continuously improve with more compute and longer thinking sessions does not hold up under controlled testing. Current large reasoning models show genuine advantages in specific complexity ranges, but they hit a hard wall when problems become sufficiently difficult. Rather than treating reasoning models as general problem solvers, the research suggests viewing them as specialized tools with well-defined limits. Future progress will likely depend on new training approaches, better evaluation methods, and a clearer understanding of what it means for a machine to reason.