Reinforcement learning can exploit a model's internal feedback to improve reasoning without ground‑truth answers or external verifiers. During training, confidence and correctness often diverge; optimizing average self‑certainty alone raises confidence while harming reasoning accuracy. We introduce SIF, a method that merges comparative reward construction with selective policy updates. SIF ranks token‑level self‑certainty profiles across multiple windows and peer samples, granting positive rewards only when both voting levels supply sufficient support. Sample‑wise and batch‑wise filtering then decides which responses contribute to policy optimization. Experiments on three backbone models across twelve mathematical and scientific reasoning benchmarks show that SIF's final checkpoints achieve higher average accuracy than the best‑validation checkpoints of Intuitor and Vigor, with a 4.71 percentage‑point gain over Intuitor on Qwen2.5‑Math‑1.5B. Further analysis reveals better retention of accuracy during continued optimization, and test‑time training demonstrates sustained improvements with as few as four rollouts per prompt.
Review: SIF's dual‑layer voting and filtering effectively mitigate the confidence‑correctness gap, offering a robust route for self‑supervised reasoning driven by internal feedback.