DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses

One-line takeaway

DiscoveryRSI defines valuable, testable hypotheses, grows hypothesis trees via multi-agent evolutionary search, compiles chosen investigations into executable tasks, and recursively learns from them, e.g. via hypothesis credit optimization.

DiscoveryRSI 先给出“有价值、可检验假设”的操作定义，再用多 agent 进化搜索长出假设树，把选中的研究编译成可执行任务，并通过假设功劳分配等机制递归学习如何发现假设。