The Coach That Trains AI to Think Like a Scientist
Coach → Train → Review → Improve → Excel
The philosophy behind MarkScientist
Like a coach defines what excellence looks like, MarkScientist establishes scientific quality standards for AI research agents.
Continuously evaluates research output, identifying strengths and areas for improvement with detailed feedback.
Pushes research agents to iterate and improve through review-driven loops until quality thresholds are met.
Cultivates scientific taste over time through calibrated feedback, helping agents internalize what makes research excellent.
How MarkScientist coaches AI research agents to excellence
Three-agent workflow (Proposer, Solver, Reviewer) creates a structured training regimen for systematic research skill development.
JudgeBuddy system provides 12 specialized reviewer personas — like assistant coaches, each focusing on different aspects of research quality.
Review-driven iteration loops act like practice drills — repeat until performance meets the coach's quality standards.
Taste learning calibrates standards through feedback, helping agents develop the intuition of an experienced researcher.
How the coach guides research agents through each phase
The "game plan" phase. Generates research proposals with hypothesis, scope, and evaluation criteria. The coach reviews the strategy before the team executes.
The "execution" phase. Carries out the approved proposal using ResearchHarness. Like athletes performing in competition, producing tangible results.
The "post-game analysis" phase. JudgeBuddy coaching staff evaluates performance with scores, critiques, and verdicts (Accept/Revise/Repropose).
Iterates until the final score meets the coach's standards. The research is now polished and ready for top-venue submission.
Specialized assistant coaches, each with unique expertise
The coach adapts training to each research phase
Get your AI research coach up and running
git clone https://github.com/InternScience/MarkScientist.git
cd MarkScientist
git submodule update --init --recursive
pip install -e .
markscientist --workflow "Write a literature review on RL for robotics"