Cat facts broke the math: one irrelevant sentence triples reasoning model errors
Ourtake
Turns out the world's smartest chatbots can be derailed by a cat fact. Sleep well, reasoning models.
What happened
A research team published a study introducing CatAttack, an automated system that uses GPT-4o and DeepSeek V3 to generate distraction sentences such as cat trivia and financial tips. These triggers tripled DeepSeek R1's error rate from 1.5 to 4.5 percent and increased token usage, with the researchers warning of risks in finance, law and healthcare.