Verified Self-Improvement (VSI) Prevents Reasoning Drift in AI Training

A new paper on arXiv (2603.21558v2) introduces Verified Self-Improvement (VSI), a method to prevent reasoning errors from compounding during self-improvement training of AI models. The researchers identify that standard filtering based solely on final answer correctness allows lucky guesses with flawed reasoning to contaminate training data, causing accuracy to stall or degrade over multiple rounds. VSI conditions data retention on step-level structural integrity by recomputing arithmetic steps via the computer-algebra library sympy, checking intermediate consistency, and enforcing domain constraints. The framework was evaluated on GSM8K using Qwen3-4B-Thinking across five rounds of self-improvement, compared against four baselines: no verification, outcome verification, majority voting, and VSI. The approach addresses a pervasive failure mode in self-improvement training where compounding reasoning errors cause accuracy to stall or degrade. By verifying the reasoning process rather than just the final answer, VSI aims to sustain capability gains across multiple training rounds.