Research map / Architectures and efficiency / Attention and transformers
Separating Representation from Reconstruction Enables Scalable Text Encoders
Megi Dervishi, Mathurin VIDEAU, Yann LeCun
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 |
|---|---|---|---|
| Scale BERT model size under frozen evaluation | BERT (flat MLM) · frozen GLUE/MTEB(eng, v2) | frozen probe score: largest BERTs fall below the smallest model (10x smaller) | Hurt |
| Separate representation from reconstruction with a cross-attention predictor (CrossBERT) (ablation) | BERT 239M · GLUE linear probe | average score: 67.4 -> 73.8 | Improved |
| Use CrossBERT architecture (ablation) | BERT 239M · MTEB(eng, v1) frozen backbone contrastive finetuning | average score: 44.3 -> 53.7 | Improved |
| Use CrossBERT architecture (ablation) | BERT 239M · GLUE full finetuning | average score: 86.4 -> 86.6 | No clear effect |
| Use CrossBERT architecture (ablation) | BERT 239M · MTEB(eng, v1) full contrastive finetuning | average score: 54.1 -> 54.5 | Improved |
| Use CrossBERT architecture (ablation) | BERT 239M · MTEB(eng, v1) frozen backbone Retrieval | Retrieval score: 19.2 -> 40.7 | Improved |
| Increase masking ratio from 20% to 50% in CrossBERT (ablation) | CrossBERT · frozen GLUE linear probe | average score: 73.8 -> 73.1 | No clear effect |
| Increase masking ratio from 50% to 65% in CrossBERT (ablation) | CrossBERT · frozen GLUE linear probe | average score: 73.1 -> 72.4 | Hurt |
| Add Complementary Masking Strategy (CMS) (ablation) | CrossBERT · sample efficiency (Appendix B) | data needed for same performance: requires half the training data | Improved |
| Probe earlier layers instead of final layer of BERT (ablation) | BERT 239M · frozen GLUE linear probe | average score: 67.4 -> 70.2 | Improved |
| Use CrossBERT's masked-token dropping in the encoder | CrossBERT · training throughput on H100 | tokens per second: 123k -> 207k | Improved |
| Warm-start contrastive adapter with the pretrained predictor (ablation) | CrossBERT frozen-backbone finetuning · MTEB | MTEB score: approx +2% | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Electra — ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators (ICLR 2020) · better on 2 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ShareBERT: Embeddings Are Capable of Learning Hidden Layers (AAAI 2024 · similarity 0.86)
- Rep2Text: Decoding Full Text from a Single LLM Token Representation (NeurIPS 2026 · similarity 0.86)
- Representation Deficiency in Masked Language Modeling (ICLR 2024 · similarity 0.86)
- When the Same Coefficients Reach Different Places: Asymmetric Realizability in Transplanting Tokenizers across Large Language Models (NeurIPS 2026 · similarity 0.85)
- ARC-Encoder: learning compressed text representations for large language models (NeurIPS 2026 · similarity 0.85)
- ExLM: Rethinking the Impact of $\texttt{[MASK]}$ Tokens in Masked Language Models (ICML 2025 · similarity 0.85)
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.
