Reproduce “Verbalizing Multi-Token Concepts in 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
Xijie Gong, Zimeng Huang, Tonghan Wang
Concept Lens reads multi-token concepts from intermediate model activations by proposing phrases from token clues, then scoring representations of those phrases against the inspected activation. Across five models and 2,400 directed clozes, average Rank@10 is 36.6% for J-lens and 54.5% for R-lens, versus 21.7% for Template Lens; target coverage remains a major bottleneck, and cloze labels may not reflect computation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.