Research map / Architectures and efficiency / Quantization
BiT: Robustly Binarized Multi-distilled Transformer
Zechun Liu, Barlas Oguz, Aasish Pappu, Lin Xiao, Scott Yih, Meng Li, Raghuraman Krishnamoorthi, Yashar Mehdad
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 |
|---|---|---|---|
| remove attention-score distillation (keep logits and block outputs only) (ablation) | binary BERT baseline · GLUE avg | avg score: 45.7 -> 48.0 | Improved |
| add two-set binarization (non-negative activations to {0,1}) (ablation) | binary BERT baseline · GLUE avg | avg score: 48.0 -> 55.3 | Improved |
| add elastic binarization with learned scale and threshold (ablation) | BiT strong baseline · GLUE avg | avg score: 55.3 -> 71.0 | Improved |
| add multi-distillation W32A32 -> W1A2 -> W1A1 (ablation) | BiT W1A1 · GLUE avg | avg score: 71.0 -> 73.5 | Improved |
| learn scale alpha in binarization (ablation) | two-set strong baseline · GLUE avg | avg score: 55.3 -> 69.2 | Improved |
| additionally learn bias/threshold beta (ablation) | BiT with learnable scale · GLUE avg | avg score: 69.2 -> 71.0 | Improved |
| binarize attention after softmax to {0,1} instead of removing softmax (Bi-Attention) (ablation) | binary BERT · GLUE avg | avg score: 51.3 -> 53.0 | Improved |
| binarize ReLU outputs to {0,1} instead of {-1,1} (ablation) | binary BERT · GLUE avg | avg score: 53.0 -> 55.3 | Improved |
| quantized teacher with fixed student init vs progressive distillation with full-precision teacher (ablation) | BiT · GLUE avg | avg score: 73.4 -> 73.5 | Improved |
| three-step schedule W1A4 -> W1A2 -> W1A1 instead of two-step (ablation) | BiT · MNLI-m | accuracy | No clear effect |
| add data augmentation for small GLUE datasets | BiT W1A1 · GLUE avg | avg score: 73.5 -> 78.0 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BiBERT 1-1-1 — BiBERT: Accurate Fully Binarized BERT (ICLR 2022) · better on 3 of 3 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
- Binarized Neural Machine Translation (NeurIPS 2023 · similarity 0.88)
- BiPFT: Binary Pre-trained Foundation Transformer with Low-Rank Estimation of Binarization Residual Polynomials (AAAI 2024 · similarity 0.87)
- FrameQuant: Flexible Low-Bit Quantization for Transformers (ICML 2024 · similarity 0.86)
- The Quantization Benefits of Residual-Free Transformers (NeurIPS 2026 · similarity 0.85)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.85)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (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.
