mlsanityData to 14 Oct 2026

CascadeBERT: Accelerating Inference of Pre-trained Language Models via Calibrated Complete Models Cascade

arXiv 2012.14682 ↗First posted 29 Dec 2020cs.CLcited by 47

Not on the map. Why?

Abstract (arXiv)

Dynamic early exiting aims to accelerate the inference of pre-trained language models (PLMs) by emitting predictions in internal layers without passing through the entire model. In this paper, we empirically analyze the working mechanism of dynamic early exiting and find that it faces a performance bottleneck under high speed-up ratios. On one hand, the PLMs' representations in shallow layers lack high-level semantic information and thus are not sufficient for accurate predictions. On the other hand, the exiting decisions made by internal classifiers are unreliable, leading to wrongly emitted early predictions. We instead propose a new framework for accelerating the inference of PLMs, CascadeBERT, which dynamically selects proper-sized and complete models in a cascading manner, providing comprehensive representations for predictions. We further devise a difficulty-aware objective, encouraging the model to output the class probability that reflects the real difficulty of each instance for a more reliable cascading mechanism. Experimental results show that CascadeBERT can achieve an overall 15% improvement under 4× speed-up compared with existing dynamic early exiting methods on six classification tasks, yielding more calibrated and accurate predictions.

Key facts. Cited by 47 papers in the map (data through 14 Oct 2026). First posted on arXiv 29 Dec 2020.

Builds on

No reference list yet.

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2021: 2; 2022: 9; 2023: 13; 2024: 13; 2025: 6; 2026: 4.

47 papers.

Sort:

By area:

  1. 2022Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Spacecited by 532
  2. 2022Confident Adaptive Language Modelingcited by 310
  3. 2021MoEfication: Transformer Feed-forward Layers are Mixtures of Expertscited by 152
  4. 2024LLM Inference Unveiled: Survey and Roofline Model Insightscited by 152
  5. 2023Towards Efficient Generative Large Language Model Serving: A Survey from Algorithms to Systemscited by 102
  6. 2024Not All Layers of LLMs Are Necessary During Inferencecited by 96
  7. 2023SkipDecode: Autoregressive Skip Decoding with Batching and Caching for Efficient LLM Inferencecited by 89
  8. 2024A Unified Approach to Routing and Cascading for LLMscited by 80
  9. 2021Towards Efficient NLP: A Standard Evaluation and A Strong Baselinecited by 44
  10. 2022Model Cascading: Towards Jointly Improving Efficiency and Accuracy of NLP Systemscited by 44
  11. 2022You Need Multiple Exiting: Dynamic Early Exiting for Accelerating Unified Vision Language Modelcited by 43
  12. 2022A Simple Hash-Based Early Exiting Approach For Language Understanding and Generationcited by 32
35 more
  1. 2025System-1.5 Reasoning: Traversal in Language and Latent Spaces with Dynamic Shortcutscited by 27
  2. 2022A Survey on Dynamic Neural Networks for Natural Language Processingcited by 24
  3. 2024FFN-SkipLLM: A Hidden Gem for Autoregressive Decoding with Adaptive Feed Forward Skippingcited by 23
  4. 2023Post-Abstention: Towards Reliably Re-Attempting the Abstained Instances in QAcited by 20
  5. 2024Online Cascade Learning for Efficient Inference over Streamscited by 20
  6. 2023Efficient Prompting via Dynamic In-Context Learningcited by 18
  7. 2024Agreement-Based Cascading for Efficient Inferencecited by 16
  8. 2022TangoBERT: Reducing Inference Cost by using Cascaded Architecturecited by 13
  9. 2023Alternating Updates for Efficient Transformerscited by 13
  10. 2023ConsistentEE: A Consistent and Hardness-Guided Early Exiting Method for Accelerating Language Models Inferencecited by 13
  11. 2022BabyBear: Cheap inference triage for expensive language modelscited by 10
  12. 2024When Neural Code Completion Models Size up the Situation: Attaining Cheaper and Faster Completion through Dynamic Model Inferencecited by 10
  13. 2024Can LLMs get help from other LLMs without revealing private information?cited by 9
  14. 2023Communication Efficient Federated Learning for Multilingual Neural Machine Translation with Adaptercited by 8
  15. 2023Towards an On-device Agent for Text Rewritingcited by 7
  16. 2024Hierarchical Skip Decoding for Efficient Autoregressive Text Generationcited by 7
  17. 2023Finding the SWEET Spot: Analysis and Improvement of Adaptive Inference in Low Resource Settingscited by 5
  18. 2023Bridging the Gap between Decision and Logits in Decision-based Knowledge Distillation for Pre-trained Language Modelscited by 5
  19. 2024DE3-BERT: Distance-Enhanced Early Exiting for BERT based on Prototypical Networkscited by 3
  20. 2025DYNAMAX: Dynamic computing for Transformers and Mamba based architecturescited by 3
  21. 2022Distributional Correlation--Aware Knowledge Distillation for Stock Trading Volume Predictioncited by 2
  22. 2025Position-Aware Depth Decay Decoding (D3): Boosting Large Language Model Inference Efficiencycited by 2
  23. 2025Improving Prediction Certainty Estimation for Reliable Early Exiting via Null Space Projectioncited by 2
  24. 2025Inference-Time Distillation: Cost-Efficient Agents Without Fine-Tuning or Manual Prompt Engineeringcited by 2
  25. 2023A Unified Evaluation Framework for Novelty Detection and Accommodation in NLP with an Instantiation in Authorship Attributioncited by 1
  26. 2024COSEE: Consistency-Oriented Signal-Based Early Exiting via Calibrated Sample Weighting Mechanismcited by 1
  27. 2025Kad: A Framework for Proxy-based Test-time Alignment with Knapsack Approximation Deferralcited by 1
  28. 2023Making Pre-trained Language Models both Task-solvers and Self-calibratorscited by 0
  29. 2023DPBERT: Efficient Inference for BERT based on Dynamic Planningcited by 0
  30. 2024C3: Confidence Calibration Model Cascade for Inference-Efficient Cross-Lingual Natural Language Understandingcited by 0
  31. 2024FiRST: Finetuning Router-Selective Transformers for Input-Adaptive Latency Reductioncited by 0
  32. 2026When Efficiency Backfires: Cascading LLMs Trigger Cascade Failure under Adversarial Attackcited by 0
  33. 2026t-gems: text-guided exit modules for decreasing clip image encodercited by 0
  34. 2026AutoRelAnnotator: Calibrated Model Cascades for Cost-Efficient Relevance Evaluation in Sponsored Searchcited by 0
  35. 2026Structure, Association, and Decision Value: Representation-Based Difficulty Estimation for Adaptive Inference in African-Language NLIcited by 0