Scaling Scientific Discovery Environments for Turn-Level Agentic RL
What this work does
Constructs data-analysis tasks with hidden evidence DAGs, synthesizes trajectories aligned with evidence primitives, and uses verified potential increments for turn-level RL.
Connection to discovery
A close precedent for collecting training trajectories and crediting intermediate scientific evidence.
Evidence
The body specifies frontier-node acceptance, repeated-work rejection, same-turn baselines and SFT-versus-GRPO-versus-DiscoPO comparisons.
Scope and limitations
The evidence graph and allowed primitives are supplied by task construction. Enforcing one frontier primitive per accepted action can favor a particular workflow; this is not new physical experimentation or unrestricted mechanism discovery.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.