training problem

[JinLong-Qin]

Hello, thank you very much for sharing this excellent work! I have two small questions regarding the evaluation, and I apologize if I overlooked something in the paper or the code:

• In the evaluation metrics, the calculation of form energy under Property seems to be missing. Could you kindly clarify how the form energy is specifically calculated in the paper?

• I followed the README instructions and used the default settings in the code for training. However, the evaluation results I obtained show a noticeable gap compared with the results reported in the paper. May I ask whether there are any additional tricks or design choices applied during training that might not have been explicitly mentioned?

Thank you very much for your time and kind help! I hope to better reproduce the results of the paper with your guidance.

[kdmsit]

Thanks for reaching out and appreciating our work.

Regarding the calculation of formation energy under Property: We followed the implementation provided in CDVAE. However, this metric has recently become obsolete, and most recent papers hardly report it anymore. Instead, the focus has shifted toward stability. For this reason, we did not include this part in the GitHub code.

On the compositional validity results: It’s a bit surprising for me as well that you obtained even higher compositional validity (which is close to state-of-the-art for MP-20). As far as I recall, we ran experiments with different seed values and reported the best results. You may want to try varying seeds too. Also, note that property statistics can vary quite a lot, and you may observe different values across different runs.