Reproduce “CulTrace: Tracing Internal Cultural Reasoning in Large 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

Haeun Yu, Arnav Arora, Seogyeong Jeong, Nadav Borenstein, Siddhesh Pawar, Jisu Shin, Jiho Jin, Junho Myung and 2 more

CulTrace trains layer-specific adapters for a natural-language decoder to describe activations during cultural question answering, making changes across layers inspectable. Across two datasets, descriptions show a broad-to-specific pattern: topic engagement, culture resolution, then an answer. In Llama-8b on BLEnD, target-culture labels first appear at mean layers 12.8 for China and 21.9 for Algeria; this measures timing, not accuracy or cause.

Summary by alphaXiv.

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