Reproduce “Explaining How Transformers Use Context to Build Predictions” 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

Javier Ferrando, Gerard I. Gállego, Ioannis Tsiamas, Marta R. Costa-jussà

The authors introduce Logit and ALTI-Logit, methods that trace layer-wise updates to a next-token logit back to input words. For “A report about the Impressionists has,” the analysis assigns a role to “report” in favoring “has” over “have.” On GPT-2 XL’s BLiMP darn subset, ALTI-Logit reaches MRR 0.945 versus Logit’s 0.873, an evidence-alignment result, not causal proof or general accuracy.

Summary by alphaXiv.

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