Reproduce “Explaining and interrupting collective misalignment in language model systems” 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
Edoardo Chidichimo
CDA identifies which messages or groups support a specified decision and tests removals by replaying the deciding agent. In 220 scheduling cases, false answers fell from 58% with the full conversation to 2% after CDA-selected removals, versus 50% after removing the most necessary message. These task-specific replay results do not prove that each removal caused or generally corrected the belief.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.