Repair, Retry, and the Cost of Hindsight: Verifier Feedback in Formal Theorem Proving

Explain Like I'm Normal

Researchers investigated why 'self-repair' in mathematical theorem proving often fails to beat simple random guessing. The study found that while compiler feedback helps, showing the model its previous failure actually confuses it, creating a zero-sum gain. This suggests that for AGI to master formal logic, we need better ways to handle 'hindsight' than just dumping error logs into a prompt.