SVR-R1: Bootstrapping Multimodal Reasoning with Self-Verification in Reinforcement Learning

SVR-R1 presents a multi-turn reinforcement-learning framework for multimodal reasoning. For each query, the model first proposes an answer and then produces a binary self-verdict using the same weights. A “No” triggers a second-chance rethink, while a “Yes” or turn limit finalizes the response for outcome-based reward computation. The system combines GRPO with asynchronous multi-turn rollouts and requires neither external supervision nor auxiliary critics. The authors report substantial gains over standard GRPO baselines on vision-language reasoning benchmarks. Training reportedly reduces verification turns while improving accuracy, suggesting that self-correction becomes increasingly internalized.