DIVE: Embedding Compression via Self-Limiting Gradient Updates
The paper titled 'DIVE: Embedding Compression via Self-Limiting Gradient Updates' introduces a new method for compressing high-dimensional embeddings from large language models. The proposed DIVE method utilizes a self-limiting hinge-based triplet loss and a head-wise NT-Xent contrastive loss to improve retrieval performance, especially in scenarios with limited labeled data. Results show that DIVE outperforms existing compression methods across multiple datasets and compression ratios.
- ▪DIVE addresses overfitting issues in embedding compression methods when labeled data is scarce.
- ▪The method employs a self-limiting hinge-based triplet loss to control gradient updates.
- ▪DIVE outperforms three baseline adapters on six BEIR datasets at various compression ratios.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computation and Language arXiv:2605.20689 (cs) [Submitted on 20 May 2026] Title:DIVE: Embedding Compression via Self-Limiting Gradient Updates Authors:Dongfang Zhao View a PDF of the paper titled DIVE: Embedding Compression via Self-Limiting Gradient Updates, by Dongfang Zhao View PDF HTML (experimental) Abstract:High-dimensional embeddings from large language models impose significant storage and computational costs on vector search systems. Recent embedding compression methods, including Matryoshka-Adaptor (EMNLP 2024), Search-Adaptor (ACL 2024), and SMEC (EMNLP 2025), enable dimensionality reduction through lightweight residual adapters, but their training objectives cause severe overfitting when labeled data is scarce, degrading retrieval performance below the…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
