Lack of Baseline

[HowieHwong]

Great work. But it seems missing one important work: UniGen (ICLR2025, https://arxiv.org/pdf/2406.18966). There are some similar points between these two methodologies (root tag -> attribute, difficulty control mechanism, as well as self-refinement).

I cannot find any emails in arxiv paper so I have to leave the text here.

[bittersweet1999]

Hi, thank you for your reminder. I have read your paper. The difference between our methods lies in whether they rely on the initial dataset and the subsequent processing logic for refinement. However, the refinement and data filtering logic in your method are of great value. We will include UniGen as a baseline method in subsequent versions. Once again, thank you for your reminder.

BTW, we have used DeepSeek to synthesize a batch of long COT reasoning data with the help of Condor's data, and we have verified that the model's reasoning ability can be significantly improved on this kind of open synthetic data. I noticed that UniGen's method can be validated through some RAG approaches to obtain more accurate data. We are also working on this aspect, and perhaps we can collaborate to advance the synthesis of related reasoning data.