DeepSeek-R1 is Smarter—but at what cost? - Deepseek-R1 Hallucinates More Than DeepSeek-V3! DeepSeek-R1 has made waves in the AI world with its impressive reasoning capabilities and open-source release. But how does it fare in terms of hallucination rate? Our latest research at Vectara reveals a surprising 4x increase in hallucination rate compared to its predecessor, DeepSeek-V3. Using Vectara’s HHEM model and Google’s FACTS evaluation, we found that higher reasoning ability may come at the cost of faithfulness—but does it have to? Our analysis dives into whether hallucination is an inevitable tradeoff for reasoning models and how better training could mitigate this issue. 📊 Read the full breakdown by Forrest Sheng Bao (@forrestbao), Ofer Mendelevitch (@ofermend), and Chenyu Xu: @deepseek_ai