Learning Self-Correction in Vision-Language Models via Rollout Augmentation
This work addresses the problem of sparse learning signals for self-correction in vision-language models, which is incremental as it builds on existing RL methods to enhance efficiency and performance.
The paper tackled the challenge of learning self-correction in vision-language models, where sparse signals hinder reinforcement learning, by proposing a rollout augmentation framework that synthesizes dense examples and a response-masking strategy, resulting in Octopus-8B achieving state-of-the-art performance on 7 benchmarks with a 1.0 score improvement and 0.72× training time per step.
Self-correction is essential for solving complex reasoning problems in vision-language models (VLMs). However, existing reinforcement learning (RL) methods struggle to learn it, as effective self-correction behaviors emerge only rarely, making learning signals extremely sparse. To address this challenge, we propose correction-specific rollouts (Octopus), an RL rollout augmentation framework that synthesizes dense self-correction examples by recombining existing rollouts. This augmentation simultaneously improves sample efficiency due to rollout reuse and stabilizes RL optimization through balanced supervision. Furthermore, we introduce a response-masking strategy that decouples self-correction from direct reasoning, avoiding signal conflicts and enabling both behaviors to be learned effectively. Building on this, we introduce Octopus-8B, a reasoning VLM with controllable self-correction capability. Across 7 benchmarks, it achieves SoTA performance among open-source VLMs, outperforming the best RLVR baseline by 1.0 score while requiring only $0.72\times$ training time per step.