What's this about?
Agents reconstruct the behaviour of existing scientific programs from public examples, then hidden cases and domain-aware verifiers determine whether the recreation is genuinely correct.
What exactly is self-supervised reconstruction?
Running existing software automatically produces both the challenge and its correct outputs, avoiding a human-written solution for every training task.
Why does this matter?
The method scales across 500 workflows and six scientific domains while resisting superficial shortcuts. Its verified trajectories improved a smaller model on an external terminal benchmark.