偏好对齐的低秩适配合并：保持子空间覆盖与解决方向各向异性
Preference-Aligned LoRA Merging: Preserving Subspace Coverage and Addressing Directional Anisotropy
TARA-Merging 提供了一种无需额外训练即可合并多个 LoRA 模块的方法，视觉研究者可将其用于多任务模型融合、领域自适应或个性化模型集成，尤其适合需要保持各任务性能的场景。
原文依据 Merging multiple Low-Rank Adaptation (LoRA) modules is promising for constructing general-purpose systems
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论合并多个低秩适配（LoRA）模块对于构建通用系统具有前景，但由于 LoRA 更新方向跨越不同子空间且贡献不均，这一过程颇具挑战。朴素合并时，这种不匹配可能削弱对某些任务损失最关键的方向，同时过度强调相对不重要的方向，最终降低模型忠实表示所有任务的能力。我们从两个视角重新审视该问题：子空间覆盖，刻画 LoRA 方向覆盖多样表示方向的广度；各向异性，反映这些方向上影响力的不均衡。我们提出 TARA-Merging（任务-秩各向异性对齐），利用偏好加权的交叉熵伪损失对齐合并权重，同时保留任务相关的 LoRA 子空间。这通过方向级重加权确保广泛的子空间覆盖并缓解各向异性。在八个视觉和六个 NLI 基准上，TARA-Merging 一致优于原始和 LoRA 感知基线，展现出强鲁棒性和泛化性，并突显了在 LoRA 合并中同时处理子空间覆盖和各向异性的重要性。
英文原文摘要 1221 字符
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. We revisit this problem through two perspectives: subspace coverage, which captures how broadly LoRA directions cover diverse representational directions, and anisotropy, which reflects the imbalance of influence across those directions. We propose TARA-Merging (Task-Rank Anisotropy Alignment), which aligns merging weights using a preference-weighted cross-entropy pseudo-loss while preserving task-relevant LoRA subspaces. This ensures broad subspace coverage and mitigates anisotropy via direction-wise reweighting. Across eight vision and six NLI benchmarks, TARA-Merging consistently outperforms vanilla and LoRA-aware baselines, demonstrating strong robustness and generalization, and highlighting the importance of addressing both subspace coverage and anisotropy in LoRA merging.