<p>Model merging offers a training-free solution to the prohibitive storage costs of the one-model-per-task paradigm. However, current merging techniques present a difficult trade-off. Creating a single model is storage-efficient but hurts performance, while adapter-based methods perform well but introduce a new storage burden. This work introduces CF-STAR, a framework that resolves this dilemma by creating highly compressible adapters. Our approach redefines the adapter itself. Instead of representing the deviation from a pre-trained model, our adapter is defined by the deviation from the multi-task average. We call this new representation the <i>centralized</i> task vector (CTV). This CTV represents a purer form of task-specific knowledge within the merging context, making it fundamentally more compressible. CF-STAR exploits this with a novel <i>low-rank plus sparse</i> decomposition tailored to this representation, capturing both global structure and critical details. Furthermore, the entire pipeline is designed to be synergistic with low-bit quantization, further enabling extreme compression. On diverse benchmarks spanning image classification, NLP, and dense prediction, CF-STAR sets a new state of the art on the accuracy-storage Pareto frontier, achieving up to <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(40\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>40</mn> <mo>×</mo> </mrow> </math></EquationSource> </InlineEquation> adapter compression over strong baselines while maintaining competitive performance.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

CF-STAR: Highly compressible adapters for model merging via centralized task vectors

  • Jialin Wu,
  • Jian Yang,
  • Jiajun Wen,
  • Junjie Cao,
  • Zhiyong Yu

摘要

Model merging offers a training-free solution to the prohibitive storage costs of the one-model-per-task paradigm. However, current merging techniques present a difficult trade-off. Creating a single model is storage-efficient but hurts performance, while adapter-based methods perform well but introduce a new storage burden. This work introduces CF-STAR, a framework that resolves this dilemma by creating highly compressible adapters. Our approach redefines the adapter itself. Instead of representing the deviation from a pre-trained model, our adapter is defined by the deviation from the multi-task average. We call this new representation the centralized task vector (CTV). This CTV represents a purer form of task-specific knowledge within the merging context, making it fundamentally more compressible. CF-STAR exploits this with a novel low-rank plus sparse decomposition tailored to this representation, capturing both global structure and critical details. Furthermore, the entire pipeline is designed to be synergistic with low-bit quantization, further enabling extreme compression. On diverse benchmarks spanning image classification, NLP, and dense prediction, CF-STAR sets a new state of the art on the accuracy-storage Pareto frontier, achieving up to \(40\times \) 40 × adapter compression over strong baselines while maintaining competitive performance.