"Discovery Intelligence" For Scientific AI Agents

Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents (harness optimization coupled with LLM GRPO Learning from Human scientific experts). ScienceBuddy: Model–Harness Co-Evolution. ScienceBuddy converts real researcher requests, scientific artifacts, tool traces, and human corrections into executable discovery tasks and evaluation rubrics. It then co-evolves the agent’s harness and LLM weights across GWAS (Genome-Wide Association Study), literature, database, and protocol re