ScientistTwo: Pioneering the Human Knowledge Frontier with Autonomous AI

ScientistTwo treats autonomous scientific research as a multi-stage, multi-agent workflow covering idea generation, subset evaluation, full experiments, ablations, drafting, peer review, rebuttal experiments, and meta-review.

Expert-level autonomous research requires more than generating a plausible paper. A useful system must test ideas, analyze failures, run ablations, respond to criticism, and produce reproducible evidence.

The challenge is coordinating research stages while preventing weak ideas from consuming full experimental budgets and ensuring that review feedback leads to new evidence rather than cosmetic text changes.

ScientistTwo builds on agentic workflows such as ReAct and AutoGen, software-engineering agents, autonomous ML systems, AI Scientist variants, and automated peer-review systems such as ScholarPeer.

Ideas are screened on benchmark subsets before full experiments. Promising ideas proceed through refinement, ablation, paper drafting, peer review, rebuttal experiments, and final meta-review.

Specialized Idea Generator, Evaluator, Analyzer, Writer, Peer-Review, and Meta-Review agents coordinate the research lifecycle. Durable workflow state and executable experiments connect the stages.

The evaluation covers 107 research problems and reports outcomes across venues including ICLR, NeurIPS, and ICML Spotlight, with comparison against ScientistOne and ScholarPeer-based evaluation.

ScientistTwo is primarily an agentic orchestration system using language models and tools rather than a new foundation-model training procedure. Research improvement happens through execution, evaluation, review, and rebuttal loops.

Across 107 problems, ScientistTwo reports 86 successful advances and a 25.2% average relative gain. Venue-level results include ICLR 4/5, NeurIPS 33/38, and ICML Spotlight 49/64 successful advances in the reported evaluation.

ScientistOne and ScholarPeer provide comparison points, alongside configurations with different peer-review and rebuttal policies.

ScientistTwo reports higher research-advancement and review outcomes in the reported evaluation, while rebuttal rounds increase reported acceptance in the ScholarPeer analysis.

The peer-review ablation compares no rebuttal with one and two rebuttal rounds, testing whether review feedback triggers new evidence gathering.

Evaluator quality, reviewer correlation, compute budgets, reproducibility, and possible evaluation leakage constrain interpretation.

ScientistTwo treats autonomous research as a durable workflow with subset screening, ablations, executable experiments, review, rebuttal, and held-out evaluation.

ScientistTwo primary paper and related work including ReAct, AutoGen, SWE-agent, OpenHands, AIDE, MLE-STAR, AI Scientist, AI Scientist-v2, PaperOrchestra, and ScholarPeer.