Google's ScientistTwo: autonomous multi-agent research beats baseline on 86 of 107 problems
2C_ornot2C · x · 2026-11-09
ScientistTwo (arXiv 2609.19644) is a fully autonomous multi-agent framework for problem-driven scientific research: given a human expert's challenge, it establishes SOTA baselines, formulates novel hypotheses, and orchestrates specialized agents through an end-to-end discovery cycle — multi-dataset experiments, automated ablations, and a closed-loop simulated peer-review/rebuttal engine. Benchmarked against ICLR/ICML/NeurIPS papers, it reports improvements on 86 of 107 research problems.
The accompanying discussion contrasts it with GIE: ScientistTwo automates a recognizable research workflow, while GIE searches for useful representations and relational structures that must survive consequential experiments rather than merely score well on benchmarks.
Related event: Google's ScientistTwo: Multi-Agent AI Automates Scientific Research(2 posts)→
More from AGI Musings
- Anti-AI Arguments Are Arguments Against Intelligence Itself: A Four-Point Case — JHochderffer · 2026-11-09
- GitHub study: engineer productivity rises 2.8% per 10% growth of local tech cluster — lemire · 2026-11-09
- Human in the loop is the weak link: reflections on manipulation after Grok's psychology explainer — TheTuringPost · 2026-11-09
- Making up numbers transparently beats hiding behind vibes, researcher argues in EA debate — davidmanheim · 2026-11-09
- Study: AI chats introduce a new vendor in two-thirds of software buying conversations — VeryWellVersed · 2026-11-09
- Hot take: good AI judgment is now the new median reviewer — 3scorciav · 2026-11-09