MachinoAI Research
ScientistTwo: Pioneering the Human Knowledge Frontier with Autonomous AI
Automate the end-to-end research loop from problem definition through experiments, ablations, review and rebuttal.
TL;DR
ScientistTwo reports 86 successful advances across 107 problems and a 25.2% average relative gain over human SOTA baselines.
Why It Matters
It treats autonomous research as a stateful workflow with explicit evidence and review gates.
Research Brief
The shortest useful explanation.
ScientistTwo reports 86 successful advances across 107 problems and a 25.2% average relative gain over human SOTA baselines.
Core Explanation
A fully autonomous multi-agent research framework that generates ideas, runs experiments, performs ablations, drafts papers, conducts peer review and rebuttal experiments, and meta-reviews the result.
Why It Matters
It treats autonomous research as a stateful workflow with explicit evidence and review gates.
ScientistTwo overview
Original paper architecture figure.
Section 01
The Problem
01 THE PROBLEM
Expert-level autonomous research needs multi-dimensional evaluation, ablation-driven analysis and active response to peer review, not only single-metric optimization.
Section 02
The Big Idea
02 THE BIG IDEA
ScientistTwo makes the full research lifecycle an explicit multi-agent loop: limitation extraction, idea generation, subset testing, full experiments, ablations, drafting, peer review, rebuttal experiments and meta-review.
Section 03
How It Works
03 HOW IT WORKS
Ideas are screened on benchmark subsets before full experiments; promising ideas are refined, ablated, drafted, reviewed and rebutted with new experiments before finalization.
Section 04
The Math
04 THE MATH
The system maps a scientific problem to a paper and reproducible codebase through specialized agents; review scores use a 1–10 scale and acceptance thresholds.
Section 05
Architecture
05 ARCHITECTURE
Idea Generator, Evaluator, Analyzer, Writer, Peer-Review and Meta-Review groups coordinate specialized critics and refiners.
Section 06
Experiments
06 EXPERIMENTS
On 107 problems, ScientistTwo reports 86 successful advances and 25.2% average relative gain. ScientistOne: 21 papers, ScholarPeer 3.8±1.2/14.3%, Stanford 4.1±0.7/0%. ScientistTwo: 86 papers, ScholarPeer 7.5±1.3/91.9%, Stanford 5.7±0.6/72.1%. Venue generated results: ICLR 4/5 with 7.0±1.2/100.0% and 5.4±0.2/75.0%; NeurIPS 33/38 with 7.3±1.7/87.9% and 5.6±0.7/75.8%; ICML Spotlight 49/64 with 7.6±1.0/93.9% and 5.7±0.6/69.4%; overall 86/107 with 7.5±1.3/91.9% and 5.7±0.6/72.1%.
Section 07
What We Learned
07 WHAT WE LEARNED
Peer-review ablation reports ScholarPeer acceptance 46.9% without rebuttal, 79.6% after one round and 93.9% after two; Stanford acceptance 49.0%, 73.5% and 69.4%. The production lesson is to make review trigger evidence gathering rather than only text edits.
Section 08
What Came Next
08 WHAT CAME NEXT
Related work includes ReAct, AutoGen, SWE-agent, OpenHands, AIDE, MLE-STAR, AI Scientist, AI Scientist-v2, PaperOrchestra and ScholarPeer. Future work is broader autonomous discovery with stronger verification and reproducibility.
Section 09
Engineering Takeaways
ENGINEERING TAKEAWAYS
Use durable workflow state, executable experiments, subset-first screening, first-class ablations, tool-using review agents and a held-out evaluator. Failure modes include evaluator leakage, reviewer correlation, runaway compute and polished prose before validity.
Formulas
Problem-to-artifact mapping
A scientific problem is transformed into a paper and reproducible codebase by specialized agents.
Related Research
Generative Agents
An agent architecture combining memory, retrieval, reflection, and planning to simulate believable long-horizon behavior.
AI AgentsMIRA
MIRA is an autonomous medical agent evaluated in a sandboxed EHR workflow with tools for diagnosis, testing, treatment, and admission decisions.
AI AgentsAgentic Reasoning
A tool-using agent framework that extends LLM reasoning with web search, coding, and structured reasoning memory for deep research tasks.