ScienceBuddy's two-level learning

What's this about?

ScienceBuddy tests scientific problem solving in four families with a fixed-model inner research harness and an outer reinforcement-learning model update. The paper proposes a defined learning cycle, not limitless autonomy.

What exactly is the inner/outer loop?

The inner loop attempts a research task using the current model; the outer loop collects feedback and trains the next model. Keeping them distinct clarifies what actually improved.

What's the community saying?

1. Science problems can provide concrete feedback for iterative agents.
2. Reward quality matters: a wrong proxy can be gamed.
3. Transfer beyond the four tested task families remains unknown.