SVR Trains Models to Decide Their Own Stopping Point, Hitting 0.563 Macro-Average in 2.99 Turns Instead of Ten

arXiv·medium signalXBlueskyLinkedInCopy link

Self-Verifying Refinement (arXiv 2607.28457, July 30) is an oracle-free multi-turn RL framework where the model emits a solution plus a discrete correctness verdict and a confidence score each turn, keeping its answer only when the verdict is Correct and confidence clears a threshold, otherwise refining using its own self-verification. Ground-truth correctness shapes training rewards only and is never exposed at inference. Trained with GRPO on fixed-horizon trajectories, with adaptive stopping activated only at inference, Qwen3.5-2B reaches 0.563 macro-average accuracy across seven math reasoning benchmarks in 2.99 turns on average — beating standard GRPO, strong multi-turn baselines, and a fixed-budget oracle-guided score-feedback reference while using far fewer turns than fixed ten-turn inference.