Research map / Architectures and efficiency / Attention and transformers
ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators
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 |
|---|---|---|---|
| replace masked language modeling with replaced token detection | ELECTRA versus BERT-Small · GLUE dev | score: 75.1 -> 79.9 | Helped |
| train using replaced token detection rather than MLM | ELECTRA-400K versus BERT (ours), equal training FLOPs · GLUE dev | Avg.: 87.2 -> 89.0 | Helped |
| train ELECTRA-1.75M rather than ELECTRA-400K | ELECTRA · GLUE dev | Avg.: 89.0 -> 89.5 | Helped |
| train ELECTRA-1.75M rather than ELECTRA-400K | ELECTRA · SQuAD 2.0 dev | EM: 86.9 -> 88.0 | Helped |
| predict identities of all input tokens with MLM (ablation) | ELECTRA · GLUE dev | score: 82.2 -> 84.3 | Helped |
| replace masked tokens with generator samples instead of [MASK] (ablation) | BERT MLM · GLUE dev | score: 82.2 -> 82.4 | Helped |
| compute discriminator loss only on masked positions (ablation) | ELECTRA · GLUE dev | score: 85.0 -> 82.4 | Hurt |
| tie token embeddings between generator and discriminator (ablation) | ELECTRA · GLUE dev | score: 83.6 -> 84.3 | Helped |
| tie all generator and discriminator weights instead of only token embeddings (ablation) | ELECTRA · GLUE dev | score: 84.3 -> 84.4 | No clear effect |
| train the generator adversarially with reinforcement learning instead of maximum likelihood (ablation) | ELECTRA · GLUE downstream tasks | score | Hurt |
| train the generator adversarially instead of with maximum likelihood (ablation) | ELECTRA generator · Masked language modeling | accuracy: 65% -> 58% | Hurt |
| use a two-stage generator-then-discriminator training procedure instead of joint training (ablation) | ELECTRA · GLUE downstream tasks | score | Hurt |
Closest papers by meaning
- Representation Deficiency in Masked Language Modeling (ICLR 2024 · similarity 0.91)
- ExLM: Rethinking the Impact of $\texttt{[MASK]}$ Tokens in Masked Language Models (ICML 2025 · similarity 0.88)
- DEPT: Decoupled Embeddings for Pre-training Language Models (ICLR 2025 · similarity 0.88)
- Think before you speak: Training Language Models With Pause Tokens (ICLR 2024 · similarity 0.87)
- Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models (ICLR 2025 · similarity 0.87)
- On the Generalization Ability of Next-Token-Prediction Pretraining (ICML 2025 · similarity 0.87)
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.