Reproduce “Surfing the Activation Wave: Analyzing the Information and Geometry of LLM Hidden States for Speculative Decoding” 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
Raghavv Goel, Sudhanshu Agrawal, Mukul Gagrani, Risheek Garrepalli, Mingu Lee
ELSI selects target-model layers for speculative decoding using calibration-based entropy scores and, in ELSI-C, CKA transitions, rather than fixed positions. It uses beam search to evaluate promising layer sets. On Qwen3-4B-Instruct-2507, three-layer ELSI averages 4.29 accepted tokens per forward pass versus 4.20 for EAGLE3 in one reported setting; gains vary by model and do not establish throughput improvement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.