菜单
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill adaptive-lora-personalized-ranks命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Adapt general LLMs to specialized manufacturing domains via domain-adaptive pretraining on open-access journals and visual instruction tuning. Extract 50M tokens and 24K images from peer-reviewed papers, achieve >90% accuracy on domain knowledge tasks, and enable real-time defect identification from manufacturing images.
基于 SOC 职业分类
正在显示 SKILL.md
| name | adaptive-lora-personalized-ranks |
| title | Not All Layers Are Created Equal: Adaptive LoRA Rank Allocation for Personalized Image Generation |
| version | 0.0.3 |
| engine | skillxiv-v0.0.3-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2603.21884 |
| keywords | ["Low-Rank Adaptation","Adaptive Ranks","Diffusion Models","Parameter Efficiency","Personalization"] |
| description | Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment. |
Current LoRA practice applies identical ranks uniformly across all layers, which is suboptimal because layer importance varies and subjects have different complexity. Simple subjects waste capacity with high-rank components, while complex subjects suffer from insufficient expressiveness with uniform low ranks.
The Modification: Replace fixed-rank LoRA with learnable per-layer rank parameters (νℓ) that are optimized during fine-tuning.
Technical Mechanism:
The method introduces learnable parameters νℓ for each LoRA component that control effective rank. A discretized exponential distribution imposes importance ordering on rank indices, preventing all ranks from collapsing identically.
# Variational rank framework: learnable importance weights per layer
# For each layer l, learn importance parameters nu_l that gate
# which rank dimensions activate during training
# Rank is dynamically added/removed via gated forward passes
# Forward pass with adaptive rank masking:
# out = (A @ diag(Lambda_l) @ B @ x)
# where Lambda_l is dynamically scaled based on learned nu_l
Weight rescaling through diagonal matrices (Λℓ) normalizes magnitudes during forward passes. The training loss combines three objectives:
Memory Footprint: LoRA2 achieves comparable visual quality with 0.40 GB vs. 2.80 GB for fixed rank-512 LoRA—a 7× reduction.
Quality Metrics across 29 subjects: