DeepSeek R1の推論に関する分析、「DeepSeek-R1 exhibits higher safety vulnerabilities compared to its non-reasoning counterpart DeepSeek-V3 (DeepSeek-AI et al , 2025b). We also show that the model’s reasoning capabilities can be used to generate jailbreak attacks that successfully elicit harmful responses from safety-aligned LLMs.」、「When presented with moral or cultural questions, DeepSeek-R1 reasons for significantly longer when prompted in English than when prompted in Chinese. It also provides different responses, displaying different sets of cultural values in each language」は面白い。