Understanding the results of particle guidance

[acharkq]

Hi Gabriele,

Congrats on the nice paper published in ICLR 2024!

My question is to understand particle guidance's effect to torsional diffusion.

After reading your paper, it seems that particle guidance focuses on improving diversity. And indeed, your kernel function for conformer is defined on the difference between torsion angles.

Why does this method improve precision? In my understanding, improving diversity will trade-off precision. However, as shown in your paper, it improves COV-Precision by 10% compared to torsional diffusion.

Can you provide some explanations or pointers to help me understand?

Thank you!