Research map / Architectures and efficiency / Pruning and efficient tuning
Rethinking Embedding Coupling in Pre-trained Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Decouple input and output embedding parameters while keeping their dimensions equal | mBERT · XTREME tasks | average score: 62.3 -> 62.7 | No clear effect |
| Reduce input embedding dimension from 768 to 128 and increase output dimension from 128 to 768 (ablation) | mBERT · XTREME tasks | average score: 60.1 -> 60.2 | No clear effect |
| Increase output embedding dimension from 128 to 3072 (ablation) | mBERT with input embedding dimension 128 · XTREME tasks | average score: 56.6 -> 61.8 | Helped |
| Add 11 pre-training layers and discard them before fine-tuning instead of increasing output embedding size (ablation) | mBERT with input embedding dimension 128 · XTREME tasks | average score: 60.2 -> 58.0 | Hurt |
| Reinvest saved input embedding parameters by increasing Transformer hidden size (ablation) | mBERT with input embedding dimension 128 · XTREME tasks | average score: 62.3 -> 64.5 | Helped |
| Reinvest saved input embedding parameters by increasing Transformer layer count (ablation) | mBERT with input embedding dimension 128 · XTREME tasks | average score: 62.3 -> 66.2 | Helped |
| Use RemBERT rather than XLM-R | Multilingual XTREME models fine-tuned only on English task data · XTREME tasks | average score: 71.4 -> 75.4 | Helped |
| Increase output embedding dimension from 128 to 3072 (ablation) | mBERT with input embedding dimension 128 · XTREME tasks | MLM accuracy: 59.0 -> 62.3 | Helped |
| Increase output embedding dimension (ablation) | mBERT Transformer representations · English-to-German nearest-neighbor translation | translation accuracy | Helped |
Closest papers by meaning
- Leviathan: Decoupling Input and Output Representations in Language Models (NeurIPS 2026 · similarity 0.91)
- EmbedLLM: Learning Compact Representations of Large Language Models (ICLR 2025 · similarity 0.91)
- DEPT: Decoupled Embeddings for Pre-training Language Models (ICLR 2025 · similarity 0.91)
- Scaling Embedding Layers in Language Models (NeurIPS 2025 · similarity 0.90)
- Layer by Layer: Uncovering Hidden Representations in Language Models (ICML 2025 · similarity 0.90)
- Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.