DeepSeek-R1-Zero, while demonstrating significant advancements in reasoning capabilities through reinforcement learning, has several notable limitations:
**Language Mixing and Coherence Issues: One of the primary challenges is the model's tendency to produce outputs that mix languages, particularly English and Chinese. This results in responses that can be incoherent or difficult to understand, especially when the input query is in a single language[1][4].
**Poor Readability: The readability of the model's outputs is often compromised. Users have reported issues with formatting and clarity, making it hard to follow the reasoning process or the final answers provided by the model[3][5].
**Lack of Supervised Fine-Tuning: DeepSeek-R1-Zero was developed without any initial supervised fine-tuning, which limits its ability to generate fully polished and human-aligned outputs. This absence of refinement can lead to less reliable responses compared to models that incorporate supervised learning techniques[1][3].
**Endless Repetition: The model sometimes exhibits repetitive behavior in its responses, which can detract from the overall user experience and effectiveness of the communication[5][7].
**Performance Variability on Coding Tasks: Although DeepSeek-R1-Zero shows strong performance in reasoning tasks, it performs poorly on coding challenges, indicating that its capabilities may not be as robust across all task types[1][2].
These limitations have prompted further development efforts, leading to the creation of DeepSeek-R1, which aims to address these shortcomings by incorporating supervised techniques and improving overall performance.
Citations:[1] https://www.prompthub.us/blog/deepseek-r-1-model-overview-and-how-it-ranks-against-openais-o1
[2] https://arxiv.org/html/2501.12948v1
[3] https://www.gocodeo.com/post/deepseek-r1-and-deepseek-r1-zero
[4] https://myedgetech.com/deepseek-r1-tr/
[5] https://huggingface.co/deepseek-ai/DeepSeek-R1-Zero
[6] https://adasci.org/mastering-llms-reasoning-capability-with-deepseek-r1/
[7] https://www.datacamp.com/blog/deepseek-r1
[8] https://arxiv.org/pdf/2501.12948.pdf