SimLens for Early Exit in Large Language Models: Eliciting Accurate Latent Predictions with One More Token

Documenti analoghi

• Label Words as Local Task Vectors in In-Context Learning
  di: Zheng, Bowen, et al.
  Pubblicazione: (2024)
• CITER: Collaborative Inference for Efficient Large Language Model Decoding with Token-Level Routing
  di: Zheng, Wenhao, et al.
  Pubblicazione: (2025)
• Beyond Tokens: Semantic-Aware Speculative Decoding for Efficient Inference by Probing Internal States
  di: Dong, Ximing, et al.
  Pubblicazione: (2026)
• LFED: A Literary Fiction Evaluation Dataset for Large Language Models
  di: Yu, Linhao, et al.
  Pubblicazione: (2024)
• Layer Importance and Hallucination Analysis in Large Language Models via Enhanced Activation Variance-Sparsity
  di: Song, Zichen, et al.
  Pubblicazione: (2024)