SVSR: Self-Verification and Self-Rectification Paradigm Reshapes Reliability Standards for Multimodal Reasoning

The SVSR (Self-Verification and Self-Rectification) framework explicitly integrates self-verification and self-rectification capabilities into the reasoning process through three-stage training. The semi-online DPO training, combined with high-quality reasoning trajectories filtered by teacher VLMs, enables the model to exhibit excellent performance in both explicit and implicit reasoning scenarios, aiming to address the reliability issues of shallow reasoning in current multimodal models.

Current multimodal models have shallow reasoning problems: incomplete, inconsistent, or even incorrect reasoning processes, high vulnerability (prone to errors when test data deviates from the training distribution), and lack of metacognitive ability—unable to self-check the rationality of reasoning, detect and correct errors, often giving wrong answers with confidence.

Three-Stage Training Paradigm

1. Unified Preference Dataset: Introduce bidirectional reasoning (forward: problem → answer + reverse: answer → problem verification) to build high-quality samples embedded with self-reflection signals;
2. Cold-Start Supervised Fine-Tuning: Train the model to generate explicit structured multi-step reasoning (including self-verification steps) to establish reasoning behavior patterns;
3. Semi-Online Direct Preference Optimization: The model dynamically generates reasoning trajectories, which are filtered by teacher VLMs to select high-quality samples for training, optimizing self-verification and self-rectification behaviors.

• Capability Emergence: After explicit self-reflection training, the model also shows significant improvement in implicit reasoning scenarios (directly giving answers without generating explicit trajectories), indicating enhanced internal reasoning ability;
• Experimental Verification: Leading performance across benchmarks, including improved reasoning accuracy, enhanced generalization ability (unseen tasks/formats), and better robustness (stable in adversarial/out-of-distribution scenarios).

• Methodological Insights: Training should focus on the reasoning process (not just results), and high-quality data (bidirectional structure, teacher filtering, dynamic enhancement) is key;
• Application Prospects: Education (AI tutoring showing problem-solving thinking processes and self-correcting), scientific research (analyzing experimental images and providing reasoning processes and confidence levels), content moderation (reducing misjudgments).

• Limitations: The three-stage training process is complex (high engineering investment), and teacher VLM filtering increases computational overhead and potential biases;
• Outlook: Simplify the training process, explore efficient self-verification mechanisms, expand to more modalities and reasoning tasks.