Research map / Large language models / Language, safety and interpretability
Slice-based Learning: A Programming Model for Residual Learning in Critical Data Slices
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 slice-residual attention to the backbone model | BERT-base · CoLA | overall Matthews correlation coefficient: 57.8 ($\pm$1.3) -> 58.3 ($\pm$0.7) | Helped |
| add slice-residual attention to the backbone model | BERT-base · RTE | overall F1: 67.0 ($\pm$1.6) -> 69.5 ($\pm$0.8) | Helped |
| add slice-residual attention to the backbone model | ResNet-18 · CyDet | overall F1: 39.4 ($\pm$-5.4) -> 40.9 ($\pm$3.9) | Helped |
| combine slice-indicator outputs and prediction confidences for attention (ablation) | SBL · Synthetic four-slice dataset | overall F1: 79.3 -> 82.7 | Helped |
| replace uniform weights with full attention (ablation) | SBL · Synthetic four-slice dataset | s3 F1: 73.6 -> 89.1 | Helped |
| replace full attention with uniform weights (ablation) | SBL · Synthetic four-slice dataset | s4 F1: 66.7 -> 72.0 | Mixed |
| use indicator outputs only for reweighting (ablation) | SBL · Synthetic four-slice dataset | overall F1: 82.7 -> 78.1 | Hurt |
| use prediction confidence only for reweighting (ablation) | SBL · Synthetic four-slice dataset | overall F1: 82.7 -> 79.3 | Hurt |
| use both slice-indicator outputs and prediction confidences instead of either signal alone (ablation) | SBL · Synthetic four-slice dataset | s1 F1: 61.1 -> 66.7 | Helped |
| increase model capacity using SBL rather than hard parameter sharing | SBL vs HPS · CoLA | maximum slice lift: +12.7 -> +19.0 | Helped |
| use SBL instead of the strongest baseline for average slice lift | BERT-base · RTE | average slice lift: +4.2 -> +4.6 | Helped |
| use SBL instead of hard parameter sharing | ResNet-18 · CyDet | average slice lift: -0.7 -> +2.3 | Helped |
Closest papers by meaning
- SliceGPT: Compress Large Language Models by Deleting Rows and Columns (ICLR 2024 · similarity 0.87)
- EfficientLLaVA: Generalizable Auto-Pruning for Large Vision-language Models (CVPR 2025 · similarity 0.83)
- Structured Packing in LLM Training Improves Long Context Utilization (AAAI 2025 · similarity 0.82)
- SCISSOR: Mitigating Semantic Bias through Cluster-Aware Siamese Networks for Robust Classification (ICML 2025 · similarity 0.82)
- IPO: Interpretable Prompt Optimization for Vision-Language Models (NeurIPS 2024 · similarity 0.82)
- Demystifying CLIP Data (ICLR 2024 · similarity 0.82)
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.