
OpenAI's Fancy New Reasoning Models Are Making Stuff Up More, Not Less
Ourtake
Thinking harder, it turns out, can also mean inventing harder. OpenAI says it can't yet explain why.
What happened
OpenAI's internal testing report found its new o3 and o4-mini reasoning models hallucinate more than the earlier o1 model. On the PersonQA benchmark, o4-mini had a 0.48 hallucination rate, o3 0.33 and o1 0.16. OpenAI said more research is needed to understand the cause.