FastE: Readout-Triggered Token Compression for LLM Embedding Inference
This study identifies depth-dependent prefix redundancy in final-readout LLM embedding models, notably across representative backbones including Qwen3-Embedding and Qwen3-VL-Embedding, and introduces FastE, a training-free, plug-and-play method that offers practical value for scalable embedding generation in retrieval,...
Jin-Song Shu, Jin-Yong Wen, Bao-Kun Wang et al.
· 0 citations