Giving an AI room to think through its problems is not enough to make it as smart as a basic calculator.
April 29, 2026
Original Paper
Barriers to Universal Reasoning With Transformers (And How to Overcome Them)
arXiv · 2604.25800
The Takeaway
Most researchers believe that Chain-of-Thought reasoning allows transformers to solve any problem if given enough tokens. This paper identifies a structural limit where models fail on problems larger than their training set unless their vocabulary also grows. By introducing signpost tokens, the researchers enabled transformers to simulate a Turing machine with linear efficiency. This tweak allows the models to generalize logic to much longer and more complex sequences than previously possible. It demonstrates that the architecture itself requires specific markers to maintain logical consistency across large-scale tasks.
From the abstract
Chain-of-Thought (CoT) has been shown to empirically improve Transformers' performance, and theoretically increase their expressivity to Turing completeness. However, whether Transformers can learn to generalize to CoT traces longer than those seen during training is understudied. We use recent theoretical frameworks for Transformer length generalization and find that -- under standard positional encodings and a finite alphabet -- Transformers with CoT cannot solve problems beyond $TC^0$, i.e. t