提出TARA-Merging方法，解决LoRA合并时方向不均衡问题。
Preference-Aligned LoRA Merging: Preserving Subspace Coverage and Addressing Directional Anisotropy
- 通过偏好加权交叉熵优化合并权重，对齐不同任务方向
- 在8个视觉和6个NLI基准上均超越基线，提升泛化能力
- 适合需要多任务兼容的轻量级模型部署场景
合并多个低秩适配（LoRA）模块有助于构建通用系统，但挑战在于LoRA更新方向分布在不同子空间且贡献不均。直接合并会导致关键任务方向被弱化，次要方向被过度强调，影响模型对所有任务的表征能力。本文从子空间覆盖度和方向各向异性两个角度重新审视该问题。提出TARA-Merging（任务-秩各向异性对齐）方法，利用偏好加权交叉熵伪损失对齐合并权重，同时保留任务相关子空间。该方法确保广泛的子空间覆盖并缓解方向失衡。在8个视觉和6个自然语言推理（NLI）基准上，TARA-Merging consistently优于基线，展现出强鲁棒性和泛化能力，凸显了在合并中同时处理子空间覆盖与各向异性的重要性。
原文摘要 · Abstract (English)
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.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。