Effective classifier-free guidance at high guidance strength for MolHIT

Develop sampling algorithms or model architectures for MolHIT’s conditional discrete molecular-graph diffusion that remain effective when using classifier-free guidance with guidance strength w greater than 1, ensuring reliable conditional generation under increased guidance without degrading performance.

Background

MolHIT uses a conditional graph transformer and classifier-free guidance (CFG) for multi-property guided molecular generation. In practice, the authors set the guidance scale to w=1.0 and observe that increasing the guidance strength beyond unity did not consistently improve property alignment in their discrete graph-diffusion framework. This suggests a limitation of existing sampling or model configurations under higher CFG weights.

The authors explicitly state that designing a better sampler or model to function effectively at higher guidance strengths is left for future research, indicating an open problem in achieving stable and beneficial CFG scaling for discrete molecular graph diffusion models like MolHIT.