Research map / Architectures and efficiency / Quantization
BiBERT: Accurate Fully Binarized BERT
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 |
|---|---|---|---|
| Add Bi-Attention to the fully binarized baseline (ablation) | BiBERT · SST-2, MRPC, RTE, QQP | accuracy: 77.6 -> 82.1; 70.2 -> 70.5; 54.1 -> 55.6; 73.2 -> 74.9 | Helped |
| Add DMD to the fully binarized baseline (ablation) | BiBERT · SST-2, MRPC, RTE, QQP | accuracy: 77.6 -> 79.9; 70.2 -> 70.5; 54.1 -> 55.2; 73.2 -> 75.3 | Helped |
| Combine Bi-Attention and DMD (ablation) | Fully binarized BERT · SST-2, MRPC, RTE, QQP | accuracy: 77.6 -> 88.7; 70.2 -> 72.5; 54.1 -> 57.4; 73.2 -> 84.8 | Helped |
| Remove attention-score distillation (ablation) | Fully binarized BERT baseline · SST-2 and QQP | accuracy: +0.81%; +1.10% | Helped |
| Remove hidden-state distillation (ablation) | Fully binarized BERT baseline · SST-2 and QQP | accuracy: -0.11%; -1.62% | Hurt |
| Remove MHA-output distillation (ablation) | Fully binarized BERT baseline · SST-2 | accuracy: 4.36% drop | Hurt |
| Use 1-bit rather than 4-bit activation quantization | Quantized activations · Standard-normal directional-error simulation | optimization direction error proportion: 3.30% -> 14.36% | Hurt |
| Use BiBERT instead of the fully binarized baseline | BERT-base · GLUE dev, without data augmentation | Avg. accuracy: 50.4 -> 63.2 | Helped |
| Use BiBERT instead of the fully binarized baseline | BERT-base · GLUE dev, with data augmentation | Avg. accuracy: 52.2 -> 67.0 | Helped |
| Use BiBERT instead of BinaryBERT | BERT-base · GLUE dev, without data augmentation | Avg. accuracy: 41.0 -> 63.2 | Helped |
| Use BiBERT instead of BinaryBERT | BERT-base · GLUE dev, with data augmentation | Avg. accuracy: 50.1 -> 67.0 | Helped |
| Apply data augmentation | BiBERT · GLUE dev | Avg. accuracy: 63.2 -> 67.0 | Helped |
Closest papers by meaning
- MosaicBERT: A Bidirectional Encoder Optimized for Fast Pretraining (NeurIPS 2023 · similarity 0.88)
- ARB-LLM: Alternating Refined Binarizations for Large Language Models (ICLR 2025 · similarity 0.86)
- Mixture of Scales: Memory-Efficient Token-Adaptive Binarization for Large Language Models (NeurIPS 2024 · similarity 0.85)
- PB-LLM: Partially Binarized Large Language Models (ICLR 2024 · similarity 0.85)
- BTR: Binary Token Representations for Efficient Retrieval Augmented Language Models (ICLR 2024 · similarity 0.84)
- BiPFT: Binary Pre-trained Foundation Transformer with Low-Rank Estimation of Binarization Residual Polynomials (AAAI 2024 · similarity 0.84)
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.