What's this about?
FlowBalance lets a reasoning model learn from its own generated trajectories but calibrates that guidance against a final verifier. Guidance is kept for successful paths, reversed for failed ones and disabled when outcomes provide no preference.
What exactly is a “terminal verifier”?
It is a reliable check at the end — for example whether the final math answer is correct — that does not explain which intermediate token helped. FlowBalance uses that sparse truth to prevent dense self-feedback from rewarding confident mistakes.
Why is this worth noting?
Self-improvement loops easily amplify false confidence. The method reports better mathematical reasoning, training stability and strategy diversity than its comparison baselines.