ScientistTwo: Pioneering the Human Knowledge Frontier with Autonomous AI

ScientistTwo autonomously pushes the frontier of human knowledge across ten-plus research domains—LLMs, robotics, neuroscience, speech, robustness, reinforcement learning, game theory, privacy, optimization, and time series—producing publication-quality papers together with fully verified codebases. Across 107 research problems drawn from recent top-tier publications, ScientistTwo improves on the human state-of-the-art in 86 of them (an 80.4% success rate), with an average relative improvement of 25.2%.

ScientistTwo runs the full research cycle on its own: it generates and evolves ideas, validates them experimentally, isolates what actually works, writes the manuscript, and then subjects that manuscript to its own peer review—looping back until the work clears a publication bar.

We score every paper with two automated reviewers—ScholarPeer and the Stanford Agentic Reviewer—and compare against both prior research agents and papers actually accepted at NeurIPS 2025, ICLR 2026, and ICML 2026, whose problem specifications and codebases serve as the benchmark tasks. ScientistTwo reaches a 91.9% acceptance rate under ScholarPeer, and is the only agent to clear the Stanford Agentic Reviewer at all—72.1% of its papers meet that bar, while every baseline scores 0%.

Strong scores mean nothing if the underlying results are not real. The CoE Integrity Audit is a post-hoc framework that checks whether a paper's claims are actually supported by its code, empirical outputs, and bibliography, via four checks: score verification (re-running the repository reproduces every reported number), specification compliance (no rule-breaking or reward hacking), reference verification (no hallucinated citations), and method–code alignment (the manuscript faithfully describes the implementation).

Three dedicated refinement agents enforce these properties: a validation filter discards rule-violating solutions right after experimentation, a search-augmented LLM grounds and corrects the bibliography, and a code audit reconciles the method section against the repository. Reproducibility is handled up front, by prompting the Coding Agent to emit self-contained, re-runnable scripts.