Researchers released the Into the Unknown benchmark on 15 Oct 2026. It hands an AI agent a science paper plus a DOE lab‑provided research direction and asks it to draft a follow‑up paper. Nine frontier agents each produced 15 drafts across fields like quantum computing and materials science. None met all the scientists’ conditions, and 114 failed an integrity check for claims or data inconsistencies. The code and data are open for others to use.
Why it matters
It shows current AI agents can generate plausible research drafts but still fall short of scientific standards, highlighting the gap for future tool development.
