Reproduce “Aligning AI Teams” 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
Siddarth Srinivasan, Morgan Jane Matthews, Jascha Sohl-Dickstein, Erik Jones
The authors compare Claude models working alone with cooperative teams on synthetic sepsis, recommender, and consulting tasks, then test assigning a safety lead and sharing a safety specification. On sepsis, Mythos Preview’s nine-agent baseline scores 59 percentage points lower in sensitivity than its solo default; the combined intervention approaches measured solo reference behavior, not general safety.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.