Distributed Direct Preference Optimization

FedDPO, staleness, lower bounds, DecDPO, and real SHP ablations
20 / 20
release gates pass
600
real SHP preference pairs
5 × 120
paper-exact client split
72
FedDPO runs
32
DecDPO runs
575.65×
no-mixing consensus penalty
Distributed DPO theorem and experiment audit