Reproduce “Large-scale Neuron Community Discovery Reveals Neuron-Sparse Behaviors in Language Models” 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

Igor Shilov, Thomas Fel, Tal Haklay, Lucius Bushnaq, Owen Lewis

The authors develop a method to find neuron groups repeatedly recruited across prompts, without specifying behavior in advance. The method scores neurons by activation and prediction sensitivity, then tests groups through interventions. In Llama-3.1-8B, groups of at most 128 neurons pass causal tests for partial contributions on 10.3% of held-out prompts; analysis is limited to the final prompt position.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.