Research map / Large language models / Language, safety and interpretability
MPNet: Masked and Permuted Pre-training for Language Understanding
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 |
|---|---|---|---|
| use MPNet instead of RoBERTa | BERT base pre-training and GLUE dev evaluation · GLUE dev | average: 86.4 -> 87.9 | Helped |
| use MPNet instead of ELECTRA | BERT base pre-training and GLUE test evaluation · GLUE test | average: 0.7 point | Helped |
| use MPNet instead of RoBERTa | BERT base pre-training and SQuAD evaluation · SQuAD v1.1 dev | F1: 91.5 -> 92.7 | Helped |
| use MPNet instead of XLNet* | pre-training on Wikipedia and BooksCorpus (16GB) · RACE test | accuracy: 66.8 -> 70.4 | Helped |
| increase MPNet pre-training corpus from 16GB to 160GB | MPNet · RACE test | accuracy: 76.1 vs. 70.4 | Helped |
| remove position compensation (ablation) | MPNet · SQuAD v1.1 dev | EM: 85.0 -> 83.0 | Hurt |
| add position compensation (ablation) | MPNet · SQuAD v1.1 dev | EM: 83.0 -> 85.0 | Helped |
| remove permutation while retaining output dependency (ablation) | MPNet · SQuAD v1.1 dev | EM: 85.0 -> 84.1 | Hurt |
| add permutation while retaining output dependency (ablation) | MPNet · SQuAD v1.1 dev | EM: 84.1 -> 85.0 | Helped |
| remove permutation and output dependency (ablation) | MPNet · SQuAD v1.1 dev | EM: 85.0 -> 82.0 | Hurt |
| add permutation and output dependency (ablation) | MPNet · SQuAD v1.1 dev | EM: 82.0 -> 85.0 | Helped |
| remove whole word masking (ablation) | MPNet · SQuAD v1.1 dev | F1: 91.4 -> 90.5 | Hurt |
Closest papers by meaning
- XLNet: Generalized Autoregressive Pretraining for Language Understanding (NeurIPS 2019 · similarity 0.90)
- ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators (ICLR 2020 · similarity 0.89)
- PMI-Masking: Principled masking of correlated spans (ICLR 2021 · similarity 0.88)
- Representation Deficiency in Masked Language Modeling (ICLR 2024 · similarity 0.86)
- DEPT: Decoupled Embeddings for Pre-training Language Models (ICLR 2025 · similarity 0.86)
- Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws (NeurIPS 2026 · similarity 0.86)
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.