Scholarlink

<p>Task-specific fine-tuning has become the de facto method for adapting large language models (LLMs) to downstream objectives, but it adds considerable computational and storage overhead and can destabilise optimisation. We introduce <i>Semantic Mixup via Spherical Interpolation</i> (SMSI), a geometry-aware data augmentation scheme that operates entirely in the latent space of a frozen encoder. SMSI first normalises the encoder outputs, thereby constraining them to lie on the unit hypersphere, and then views each label region as a locally spherical submanifold on which we synthesise new samples by Slerp-interpolating between cluster-level neighbours that share the same label. Experiments on seven emotion- and sentiment-analysis benchmarks show that SMSI achieves <i>competitive performance</i> while using roughly <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(10\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>10</mn> <mo>×</mo> </mrow> </math></EquationSource> </InlineEquation> fewer trainable parameters than full fine-tuning and matching the accuracy of parameter-efficient approaches such as LoRA. These findings demonstrate that latent space semantic interpolation can offer a lightweight alternative to task-specific adaptation.</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

Semantic spherical mixup: geometry-aware data augmentation in latent manifolds for parameter-efficient language model adaptation

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Pradeep Singh,
  • Balasubramanian Raman

摘要

Task-specific fine-tuning has become the de facto method for adapting large language models (LLMs) to downstream objectives, but it adds considerable computational and storage overhead and can destabilise optimisation. We introduce Semantic Mixup via Spherical Interpolation (SMSI), a geometry-aware data augmentation scheme that operates entirely in the latent space of a frozen encoder. SMSI first normalises the encoder outputs, thereby constraining them to lie on the unit hypersphere, and then views each label region as a locally spherical submanifold on which we synthesise new samples by Slerp-interpolating between cluster-level neighbours that share the same label. Experiments on seven emotion- and sentiment-analysis benchmarks show that SMSI achieves competitive performance while using roughly \(10\times \) 10 × fewer trainable parameters than full fine-tuning and matching the accuracy of parameter-efficient approaches such as LoRA. These findings demonstrate that latent space semantic interpolation can offer a lightweight alternative to task-specific adaptation.