← Search

Jianing He

2 accepted papers

2025

COSEE: Consistency-Oriented Signal-Based Early Exiting via Calibrated Sample Weighting Mechanism

AAAI 2025technical

Early exiting is an effective paradigm for improving the inference efficiency of pre-trained language models (PLMs) by dynamically adjusting the number of executed layers for each sample. However, in most existing works, easy and hard samples are treated equally by each classifier during training, w…

2025

Improving Prediction Certainty Estimation for Reliable Early Exiting via Null Space Projection

IJCAI 2025

Early exiting has demonstrated great potential in accelerating the inference of pre-trained language models (PLMs) by enabling easy samples to exit at shallow layers, eliminating the need for executing deeper layers. However, existing early exiting methods primarily rely on class-relevant logits to