Training environments are a major bottleneck for data‑driven scientific discovery agents. Existing collections of executable and verifiable environments are small and costly to build. To tackle this, we introduce Poseidon, an automated pipeline that generates diverse, verifiable training environments at scale from public scientific code repositories. Poseidon decomposes scientific workflows into four complementary task types: Reproduction to verify existing experiments, Subflow to extract sub‑stages of a workflow, Completion to ask the model to fill missing code or parameters, and Correction to have the model fix erroneous implementations. Each task targets a distinct research capability. The pipeline also incorporates a recursive improvement loop that iteratively refines initially failed candidates, rescuing environments that would otherwise be discarded. With the same repository budget, Poseidon extracts roughly ten times more environments than previous methods. Using Poseidon we built the Pearl collection, covering 7,110 public repositories across ten scientific disciplines and containing 97,659 executable, verifiable environments—the largest and most multidisciplinary set to date. Qwen3.5 models (9B, 27B, 35B‑A3B) trained on Pearl achieve consistent gains on six challenging scientific‑agent benchmarks. Overall, Pearl’s scale and diversity open new opportunities for training effective scientific discovery agents.
Review: Poseidon’s task decomposition and recursive refinement dramatically lower the barrier to environment creation, and Pearl’s scale is poised to accelerate the development of scientific AI agents.