Scientific Knowledge
A commons of research skills — authored by human experts, executed by AI agents.
ResearchSkills
ResearchSkills.ai is an open platform that turns real scientific workflows into reusable skills for AI agents. Instead of storing raw chat logs, it reconstructs research sessions as decision trajectories — how researchers form hypotheses, diagnose failures, choose methods, and decide when to pivot — and distills them into portable skills that agents can retrieve and execute. Contributions are extracted locally, automatically de-identified, reviewed by domain experts for scientific accuracy, and published to an open library spanning 155+ scientific subdomains under CC BY 4.0. The goal is to give AI systems not just raw capability, but the tacit research judgment that determines which experiments matter, which dead ends to avoid, and when to persist versus change direction.
Scaling Behaviors of LLM Reinforcement Learning Post-Training: An Empirical Study in Mathematical Reasoning
This paper systematically reveals how model scale, data volume, and compute jointly govern the scaling behavior of RL post-training for mathematical reasoning in large language models.
