SVSR: A Self-Verification and Self-Rectification Paradigm for Multimodal Reasoning
Authors: Zhe Qian, Nianbing Su, Zhonghua Wang, Hebei Li, Zhongxing Xu, Yueying Li, Fei Luo, Zhuohan Ouyang, Yanbiao Ma
Organizations: South China Agricultural University · Monash University · University of Glasgow · University of Electronic Science and Technology of China · University of Science and Technology of China · National University of Defense Technology · Renmin University of China · South China Normal University
Current multimodal models often suffer from shallow reasoning, leading to errors caused by incomplete or inconsistent thought processes. To address this limitation, we propose Self-Verification and Self-Rectification (SVSR), a unified framework that explicitly integrates self-verification and self-rectification into the model's reasoning pipeline, substantially improving robustness and reliability in complex visual understanding and multimodal reasoning tasks. SVSR is built on a novel three-stage training paradigm. First, we construct a high-quality unified preference dataset by refining reasoning traces from pre-trained vision-language models, incorporating both forward and backward reasoning to embed self-reflective signals. Second, we perform cold-start supervised fine-tuning on this dataset to learn structured, multi-step reasoning behaviors. Third, we apply a Semi-online Direct Preference Optimization (Semi-online DPO) process, continuously augmenting the training corpus with high-quality, model-generated reasoning traces filtered by a powerful teacher VLM. This pipeline enables the model to learn, elicit, and refine its ability to self-verify and self-rectify. Extensive experiments across diverse benchmarks demonstrate that SVSR improves reasoning accuracy and enables stronger generalization to unseen tasks and question types. Notably, once trained with explicit self-reflective reasoning, the model also exhibits improved implicit reasoning ability, outperforming strong baselines even when no explicit reasoning traces are provided. These results highlight the potential of SVSR for building more dependable, introspective, and cognitively aligned multimodal systems.
Figures & tables
Figure 1: Impact of the Self-Verification and Self-Rectification (SVSR) framework. (a) Qualitative Case Study: Qualitative Example: Demonstrates SVSR’s ability to identify and correct an initially incorrect answer (720°) in a visual math problem, ultimately producing the correct solution (1080°) through self-verification and self-rectification.. (b) Quantitative Comparison Results: Reports accuracy improvements across benchmarks (e.g., MathVista, DynaMath) as the model progresses from the base model, through Cold Start SFT, to the final SVSR model with reinforcement fine-tuning.
Figure 2: Overview of the SVSR three-stage training pipeline. Stage 1: Data Construction generates self-correction trajectories by refining initial reasoning through verification and rectification steps. Stage 2: Cold Start SFT fine-tunes the model on these structured trajectories, teaching it the format of reflective, multi-step reasoning. Stage 3: Semi-Online DPO further improves the model by incorporating preference data from model-generated outputs filtered by a teacher VLM, enabling iterative self-correction refinement.
Model
MathVista
DynaMath
MathVerse
MMstar
MMvet
AI2D
Average
Closed-Source Models
GPT-4V
–
–
–
56.0
67.5
78.2
–
GPT-4o-20240513
60.0
34.5
40.6
64.7
69.1
84.6
59.0
Claude-3.5-Sonnet
64.7
36.6
44.2
65.1
70.1
81.2
60.3
Gemini-1.5-Pro
68.7
40.7
30.1
59.1
64.0
79.1
57.0
Open-Source Models
Table 1: Model performance on multimodal benchmarks. Scores represent Pass@1 accuracy (%). The best result in each benchmark column is shown in bold .
Model
Self-verification
Self-rectification
Verification Acc
Error Recall
Incorrect to Correct
Correct to Incorrect
Qwen2.5 VL-7B-Instruct
39.87
48.18
10.15
1.22
+ Cold Start
56.12
61.92
16.81
5.31
+ RFT
69.24
69.15
23.10
2.89
Table 2: Quantitative results of self-verification and self-rectification abilities across model variants.
Figure 3: Difficulty-based assessment of model accuracy and average number of trials. The bars (left axis) show the accuracy of models across five difficulty levels, while the lines (right axis) represent the average number of verification attempts required. The SFT+RFT model demonstrates both higher accuracy and more adaptive resource allocation, using fewer trials on easier problems (Levels 1-2) while investing more verification steps on difficult problems (Levels 3-5).
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Figure 4: This figure illustrates the model’s reasoning process when applying the SVSR framework to a science question about energy flow in ecosystems. The example demonstrates the three key phases of our approach: initial reasoning, where the model first analyzes which organism gets energy from plants; self-verification, where it explicitly checks its solution (indicated by "Wait, let me recheck my solution"); and self-rectification, where it identifies and corrects its misconception about the direction of energy flow in the food web. The model ultimately arrives at the correct answer (C. Rabbit) after successfully implementing the verification-correction cycle, showcasing how SVSR enables structured, multi-step reasoning with built-in error detection and correction.
We introduce Self-Verified Reasoner (SVR-R1), a multi-turn RL framework that turns a model's own verification into a learning signal for multimodal reasoning. For each query, the model proposes an answer using the same weights, and issues a binary self-verdict (Yes/No). A 'No' triggers a second-chance rethink; a 'Yes,' or a turn cap, finalizes the output for computing the outcome-based reward. SVR-R1 is implemented with GRPO and an asynchronous multi-turn rollout framework and needs no external supervision or auxiliary critics. We evaluate SVR-R1 on vision-language reasoning benchmarks and show that it improves accuracy by a large margin over strong standard GRPO baselines. Training dynamics show decreasing reliance on verification-fewer verification turns, yet higher test accuracy-indicating that the gap between verification and generation narrows as the policy internalizes self-correction and chooses the most confident answer via our framework. SVR-R1 bridges the less explored intersection of inference-time self-refinement and RL training for VLMs, offering a simple yet effective recipe for bootstrapping multimodal reasoning. We will open-source SVR-R1 to facilitate future research in VLMs.
Vision-Language Models (VLMs) have achieved remarkable progress in integrating visual perception with language understanding. However, effective multimodal reasoning requires both accurate perception and robust reasoning, and weakness in either limits the performance of VLMs. Prior efforts to enhance reasoning often depend on high-quality chain-of-thought (CoT) data, obtained via labor-intensive human annotations, costly proprietary models, or self-training methods that overlook perception. To address these limitations, we propose a simple yet effective self-training framework called See-Think-Learn (STL). At its core, STL introduces a structured reasoning template that encourages the model to see before thinking, first extracting visual attributes in textual form, then using them to guide reasoning. The framework jointly improves perception and reasoning by having the model generate and learn from its own structured rationales in a self-training loop. Furthermore, we augment the training data with negative rationales, i.e. explanations that justify why certain answer choices are incorrect, to enhance the model's ability to distinguish between correct and misleading responses. This fosters more discriminative and robust learning. Experiments across diverse domains show that STL consistently outperforms baselines trained directly only on answers or self-generated reasoning, while qualitative analysis confirms the high quality of its rationales. STL thus provides a cost-effective solution to enhance multimodal reasoning ability of VLMs.
Sourabh Sharma, Sonam Gupta, Sadbhawna
Malaviya National Institute of Technology Jaipur · IBM Research
Vision-language models (VLMs) have achieved strong performance in multimodal reasoning, yet they remain prone to generating plausible but incorrect answers. Self-verification offers a practical way to improve answer reliability without relying on external judges, but existing methods typically depend on a single verification criterion or fixed prompt, resulting in incomplete and unstable reliability estimates. We first systematically analyze how verifier capability and prompt design affect verification performance. Our findings show that stronger verifiers provide more reliable judgments, while verification performance is highly sensitive to prompt choice, with no single prompt consistently dominating across tasks. Guided by these findings, we propose MOTIVE, a Multi-View Self-VerificatiOn wiTh RelIability-Guided SelectiVE Rethinking framework for reliable multimodal reasoning. MOTIVE evaluates each candidate answer from complementary verification perspectives and learns a correctness-aligned reliability score through correctness-grounded multi-view verification learning. During inference, this score governs an accept-or-rethink decision, allowing reliable answers to be returned directly while uncertain ones trigger history-guided rethinking. Extensive experiments across diverse multimodal benchmarks and VLM backbones demonstrate that MOTIVE consistently outperforms strong self-verification and self-correction baselines. Further results show that reliable verification improves accept-or-rethink decisions and reduces unnecessary reasoning turns, enabling more reliable and efficient self-verification without an external judge.
Ziquan Zhu, Hanruo Zhu, Si-Yuan Lu +10
University of Leicester · Nanjing University of Posts and Telecommunications · University of North Carolina at Chapel Hill +5