Complex problems break Reasoning Models
June 17, 2025
We’re teaching models to think. But not like we do.
New research from Apple shows something strange:
As tasks get harder, reasoning models don’t just fail -- they think less.
They start strong, using more tokens to solve more complex problems.
Then, right before they break down, they reduce effort.
They under-think just when they need to reason more.
Even with the correct algorithm given, the collapse still happens.
Humans stall too. But we notice. We pause. We adapt.
These models don’t. Not yet.
In humans, reasoning often scales with problem complexity.
We deliberate more when we sense difficulty, recruit working memory, and apply mental models and other frameworks.
These large reasoning models (LRMs) show the opposite.
We’ve made a leap.
But adaptive reasoning *true generalization* still looks like the next one.
#AI #Reasoning #CognitiveScience #LLM #AppleResearch