The papers, preprints, and findings produced by the stack — because a scientific AI infrastructure has to prove itself by doing science.
We don’t evaluate the stack only on benchmarks. We use it to do real research — drive it at open questions, publish results, release reproducibility artifacts. Scientific Discovery is the running record of what Science Intelligence has produced using its own Knowledge, Models, Agent, and Evaluation infrastructure.
This paper envisions AI-driven Science of Science as a new paradigm for automatically discovering large-scale research patterns, simulating scientific societies, and revealing the hidden mechanisms that drive innovation beyond the reach of traditional statistical and rule-based methods.
Agentic Reinforcement Learning (Agentic RL) reframes large language models from passive text generators into autonomous agents that learn to make decisions in dynamic, partially observable environments. This survey synthesizes over 500 recent works, systematically organizing core agentic capabilities, application domains, open-source environments, benchmarks, and frameworks to guide the development of scalable general-purpose AI agents.