Richer conditioning strategies for c-QA in BM-VAE

Develop conditioning strategies beyond attribute-average biasing for conditional quantum annealing applied to the learned Boltzmann-machine prior in BM-VAEs, in order to achieve more precise and compositional control over generated outputs, and assess their performance relative to the current approach that derives external bias fields from attribute-average encoder statistics.

Background

Conditional generation in the paper is implemented by adding external bias fields to the learned Boltzmann-machine prior and sampling with c-QA. The current implementation constructs these bias fields from attribute-average encoder outputs and demonstrates successful attribute control (e.g., adding Bangs).

The Discussion explicitly notes that enriching the conditioning strategy beyond attribute-average biasing remains an open direction, suggesting the need to design and test alternative constructions (e.g., instance-specific, compositional, or learned bias fields) to improve precision and compositionality of control.