Transformer-Architecture Convergence Bounds

Extend the convergence bounds for Personalized Federated Direct Preference Optimization beyond log-linear softmax policies to transformer architectures by controlling the relevant Fisher information spectra.

Background

The theoretical analysis of Personalized Federated Direct Preference Optimization inherits log-linear softmax policy assumptions from prior FedDPO analysis. The paper identifies extending these guarantees to transformer architectures as requiring control of Fisher information spectra, which is not provided in the current work.

References

Extending the bounds to transformer architectures requires controlling Fisher information spectra and is left for future work.

— Personalized Federated Direct Preference Optimization via LoRA Adapters  (2610.12885 - Jiang et al., 9 Oct 2026) in Section “Limitations”

A variant that transmits low-rank updates while keeping adapters private is an open direction.

— Personalized Federated Direct Preference Optimization via LoRA Adapters  (2610.12885 - Jiang et al., 9 Oct 2026) in Section “Limitations”