ScientistTwo automates end-to-end scientific research

While autonomous research agents often risk becoming high-velocity paper mills churning out ungrounded claims, ScientistTwo sets a new benchmark by coupling idea generation with sandboxed code execution, real ablation experiments, and empirical rebuttals.

• –End-to-End Autonomous Workflow: Manages idea generation, benchmark filtering, sandboxed experimentation, iterative ablation studies, paper drafting, and interactive rebuttal cycles without human intervention.
• –Empirical Rebuttal Mechanism: Instead of superficially editing draft text to satisfy reviewer critiques, the rebuttal agent designs, codes, and executes targeted follow-up experiments.
• –Rigorous Artifact Auditing: Employs the CoE Integrity Audit to guarantee 100% score reproducibility, verify method-to-code fidelity, and eliminate hallucinated citations across thousands of references.
• –Surpassing Human Baselines: Outperformed existing human state-of-the-art solutions on 80.4% of evaluated benchmarks, achieving a 91.9% acceptance rate on ScholarPeer and 72.1% on the Stanford Agentic Reviewer.
• –Paradigm Shift for AI Research: Demonstrates that scientific velocity in machine learning will increasingly scale with compute, shifting human researcher efforts toward high-level direction, problem framing, and fundamental breakthrough verification.