Research map / Architectures and efficiency / RNNs and normalization
BERT Loses Patience: Fast and Robust Inference with Early Exit
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 |
|---|---|---|---|
| Use patience-based exit rather than full-depth inference | ALBERT-base · GLUE dev macro | Macro Score: 84.4 -> 85.1 | Helped |
| Use patience-based exit rather than full-depth inference | ALBERT-base · GLUE dev | Speed-up: 1.57$\times$ | Helped |
| Use patience-based exit rather than full-depth inference | BERT-base · GLUE dev MNLI | MNLI: 84.5 -> 83.6 | Hurt |
| Use patience-based exit rather than full-depth inference | BERT-base · GLUE dev SST-2 | SST-2: 92.1 -> 92.0 | No clear effect |
| Use patience-based exit rather than full-depth inference | BERT-base · GLUE dev STS-B | STS-B: 88.9 -> 88.7 | No clear effect |
| Use patience-based exit rather than full-depth inference | ALBERT · SNLI under TextFooler | After-Attack Acc.: 5.5 -> 19.3 | Helped |
| Use patience-based exit rather than full-depth inference | ResNet-56 · CIFAR-10 | Accuracy: 91.8 -> 92.0 | Helped |
| Use patience-based exit rather than full-depth inference | ResNet-56 · CIFAR-100 | Accuracy: 68.6 -> 69.1 | Helped |
| Set patience between 5 and 8 (ablation) | PABEE with ALBERT-base · MNLI, SST-2, STS-B | Accuracy | Helped |
| Set patience too large (ablation) | PABEE with ALBERT-base · GLUE tasks | Accuracy | Hurt |
| Set patience relatively small (ablation) | PABEE with ALBERT-base · GLUE tasks | Accuracy | Hurt |
| Increase model depth from ALBERT-base to ALBERT-large while using PABEE (ablation) | PABEE · GLUE dev MNLI | MNLI: 85.1 -> 86.8 | Helped |
Closest papers by meaning
- ConsistentEE: A Consistent and Hardness-Guided Early Exiting Method for Accelerating Language Models Inference (AAAI 2024 · similarity 0.89)
- RAEE: A Robust Retrieval-Augmented Early Exit Framework for Efficient Inference (ICLR 2026 · similarity 0.89)
- PoWER-BERT: Accelerating BERT Inference via Progressive Word-vector Elimination (ICML 2020 · similarity 0.88)
- EE-LLM: Large-Scale Training and Inference of Early-Exit Large Language Models with 3D Parallelism (ICML 2024 · similarity 0.86)
- PolyPythias: Stability and Outliers across Fifty Language Model Pre-Training Runs (ICLR 2025 · similarity 0.85)
- APT: Adaptive Pruning and Tuning Pretrained Language Models for Efficient Training and Inference (ICML 2024 · 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.