Early exiting

Early exiting using confidence measures

GeneralIntroduced 200066 papers

Description

Exit whenever the model is confident enough allowing early exiting from hidden layers

Papers Using This Method

Early Stopping Tabular In-Context Learning2025-06-26Languages in Multilingual Speech Foundation Models Align Both Phonetically and Semantically2025-05-26System-1.5 Reasoning: Traversal in Language and Latent Spaces with Dynamic Shortcuts2025-05-25Performance Control in Early Exiting to Deploy Large Models at the Same Cost of Smaller Ones2024-12-26COSEE: Consistency-Oriented Signal-Based Early Exiting via Calibrated Sample Weighting Mechanism2024-12-17A Stitch in Time Saves Nine: Small VLM is a Precise Guidance for Accelerating Large VLMs2024-12-04Energy-Aware Dynamic Neural Inference2024-11-04BERTer: The Efficient One2024-07-19Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity2024-06-20Predicting Probabilities of Error to Combine Quantization and Early Exiting: QuEE2024-06-20RAEE: A Robust Retrieval-Augmented Early Exiting Framework for Efficient Inference2024-05-24A Comprehensive Survey of Accelerated Generation Techniques in Large Language Models2024-05-15Computational Efficient Width-Wise Early Exiting in Wireless Communication Systems2024-05-06Kangaroo: Lossless Self-Speculative Decoding via Double Early Exiting2024-04-29LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding2024-04-25Efficient Transformer Encoders for Mask2Former-style models2024-04-23Tiny Models are the Computational Saver for Large Models2024-03-26Understanding the Training Speedup from Sampling with Approximate Losses2024-02-10EERO: Early Exit with Reject Option for Efficient Classification with limited budget2024-02-06DE$^3$-BERT: Distance-Enhanced Early Exiting for BERT based on Prototypical Networks2024-02-03