Research map / Large language models / Instruction and fine-tuning
Should We Still Pretrain Encoders with Masked Language Modeling?
Hippolyte Gisserot-Boukhlef, Nicolas Boizard, Manuel Faysse, Duarte Miguel Alves, Emmanuel Malherbe, Andre Martins, CELINE HUDELOT, Pierre Colombo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| pretrain with MLM instead of CLM | 610M encoder · QA average | F1: 42.09 -> 62.77 | Improved |
| pretrain with MLM instead of CLM | 610M encoder · SC average | accuracy: 83.58 -> 87.00 | Improved |
| pretrain with MLM instead of CLM | 610M encoder · TC average | F1: 92.69 -> 92.21 | Hurt |
| pretrain with MLM instead of CLM | 610M encoder · IR average | NDCG@10: 76.20 -> 79.55 | Improved |
| pretrain with MLM instead of CLM | 1B encoder · IR average | NDCG@10: 78.68 -> 79.95 | Improved |
| use CLM instead of MLM early in training | 610M encoder · IR average at step 1K | NDCG@10: 20.19 -> 29.94 | Improved |
| use CLM instead of MLM early in training | 610M encoder · TC average at step 1K | F1: 83.15 -> 87.01 | Improved |
| train MLM to the end instead of CLM | 610M encoder · SC average at step 40K | accuracy: 83.48 -> 86.40 | Improved |
| start with CLM then MLM (3K+9K) vs pure MLM (ablation) | 610M encoder · SC average, 12K steps | accuracy: 83.04 -> 83.67 | Improved |
| start with CLM then MLM (5K+17K) vs pure MLM (ablation) | 610M encoder · SC average, 22K steps | accuracy: 86.07 -> 86.31 | Improved |
| allocate 75% of steps to CLM (9K+3K) (ablation) | 610M encoder · SC average, 12K steps | accuracy: 83.04 -> 82.32 | Hurt |
| pure CLM instead of pure MLM (ablation) | 610M encoder · SC average, 12K steps | accuracy: 83.04 -> 80.86 | Hurt |
Closest papers by meaning
- Representation Deficiency in Masked Language Modeling (ICLR 2024 · similarity 0.90)
- Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws (NeurIPS 2026 · similarity 0.88)
- DEPT: Decoupled Embeddings for Pre-training Language Models (ICLR 2025 · similarity 0.88)
- A Comprehensive Study of Decoder-Only LLMs for Text-to-Image Generation (CVPR 2025 · similarity 0.88)
- Scaling up Masked Diffusion Models on Text (ICLR 2025 · similarity 0.88)
- ExLM: Rethinking the Impact of $\texttt{[MASK]}$ Tokens in Masked Language Models (ICML 2025 · similarity 0.88)
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.
