Preference-Aligned LoRA Merging: Preserving Subspace Coverage and Addressing Directional Anisotropy

ArXi:2603.26299v1 Announce Type: cross Merging multiple Low-Rank Adaptation (LoRA) modules is promising for constructing general-purpose systems, yet challenging because LoRA update directions span different subspaces and contribute unevenly. When merged naively, such mismatches can weaken the directions most critical to certain task losses while overemphasizing relatively less important ones, ultimately reducing the model's ability to represent all tasks faithfully.