Reproduce “Dynamic Vocabulary Pruning in Early-Exit LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jort Vincenti, Karim Abdel Sadek, Joan Velja, Matteo Nulli, Metod Jazbec
Researchers at the University of Amsterdam developed Dynamic Vocabulary Pruning (DVP), a post-hoc method to accelerate early-exit Large Language Models by drastically reducing the computational cost of confidence estimation. DVP achieves up to a 7-fold reduction in FLOPs for confidence calculation across question answering and text summarization tasks, matching baseline performance without requiring model finetuning.
Key insights
- The token eventually predicted by the full LLM tends to appear among the top-K most likely tokens very early in the forward pass (e.g., in 95% of cases by the second layer for T5-large).
- Aggressively pruning the vocabulary for confidence estimation after initial layers is possible without significant performance degradation or hindering early-exit decisions.
- Early-exit finetuning (as used in CALM) helps ensure faster token convergence, allowing for smaller pruning parameters (p and K) and greater efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.