Research map / Architectures and efficiency / Attention and transformers
Block-Skim: Efficient Question Answering for Transformer
Yue Guan, Zhengyi Li, Zhouhan Lin, Yuhao Zhu, Jingwen Leng, Minyi Guo
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 |
|---|---|---|---|
| Jointly train the Block-Skim loss with the QA objective | BERT-base · SQuAD, HotpotQA, NewsQA, NaturalQuestions, TriviaQA, SearchQA | F1: 76.78 -> 77.36 | Improved |
| Skim predicted irrelevant blocks during inference | BERT-base · SQuAD, HotpotQA, NewsQA, NaturalQuestions, TriviaQA, SearchQA | F1: 88.32 -> 88.52; 74.39 -> 74.47; 66.57 -> 65.14; 78.85 -> 78.48; 72.61 -> 72.80; 79.93 -> 79.84 | Mixed |
| Skim predicted irrelevant blocks during inference | BERT-base · Six QA datasets | Speedup: 3.01×; 2.28×; 2.53×; 2.56×; 1.81×; 3.17× | Improved |
| Freeze the Transformer and train only the Block-Skim modules instead of joint training (ablation) | Block-Skim · SQuAD | F1: 88.92 -> 88.32 | Hurt |
| Skim blocks during joint QA training (ablation) | Block-Skim · SQuAD | F1: 88.92 -> 86.83 | Hurt |
| Use block size 1 instead of 32 (ablation) | Block-Skim · SQuAD | F1: 88.92 -> 88.60 | Hurt |
| Use block size 8 instead of 32 (ablation) | Block-Skim · SQuAD | F1: 88.92 -> 88.63 | Hurt |
| Use block size 16 instead of 32 (ablation) | Block-Skim · SQuAD | F1: 88.92 -> 88.75 | Hurt |
| Use block size 64 instead of 32 (ablation) | Block-Skim · SQuAD | F1: 88.92 -> 88.65 | Hurt |
| Use block size 128 instead of 32 (ablation) | Block-Skim · SQuAD | F1: 88.92 -> 88.33 | Hurt |
| Add supporting-fact evidence labels to the skim predictor objective (ablation) | Block-Skim · HotpotQA | F1: 74.88 -> 74.85 | No clear effect |
Closest papers by meaning
- Efficient Large Language Model Inference with Neural Block Linearization (NeurIPS 2025 · similarity 0.87)
- Transformer Block Coupling and its Correlation with Generalization in LLMs (ICLR 2025 · similarity 0.85)
- Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models (ICML 2024 · similarity 0.85)
- Block Transformer: Global-to-Local Language Modeling for Fast Inference (NeurIPS 2024 · similarity 0.84)
- Beyond Attention: Breaking the Limits of Transformer Context Length with Recurrent Memory (AAAI 2024 · similarity 0.84)
- On the Robustness of Transformers against Context Hijacking for Linear Classification (NeurIPS 2025 · 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.
