Research map / Vision-language and multimodal / Multimodal LLMs
ReCoT: Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models
Mengxue Qu, Yibo Hu, Kunyang Han, Yunchao Wei, Yao Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add reflective fine-tuning (RFT) (ablation) | Formatted SFT, Mulberry-2B · ScienceQA | accuracy: 85.03 -> 86.51 | Improved |
| add reflective fine-tuning (RFT) (ablation) | Formatted SFT, Mulberry-2B · MME | score: 1869.00 -> 1931.70 | Improved |
| add reflective fine-tuning (RFT) (ablation) | Formatted SFT, Mulberry-2B · MMStar | accuracy: 46.40 -> 49.40 | Improved |
| add Consist-DPO (ablation) | SFT+RFT, Mulberry-2B · ScienceQA | accuracy: 86.51 -> 86.37 | No clear effect |
| add Consist-DPO (ablation) | SFT+RFT, Mulberry-2B · MME | score: 1931.70 -> 2070.58 | Improved |
| add Consist-DPO (ablation) | SFT+RFT, Mulberry-2B · MMStar | accuracy: 49.40 -> 50.20 | Improved |
| add Consist-DPO (ablation) | Mulberry-2B · RCS validation | RCS: 89.83 -> 97.88 | Improved |
| add Consist-DPO (ablation) | REVERIE-7B · RCS validation | RCS: 87.81 -> 91.04 | Improved |
| use 10k DPO samples instead of all ~100k (ablation) | Consist-DPO, Mulberry-2B · ScienceQA | accuracy: 81.80 -> 86.37 | Improved |
| use 10k DPO samples instead of all ~100k (ablation) | Consist-DPO, Mulberry-2B · MME | score: 1909.99 -> 2070.58 | Improved |
| use 10k DPO samples instead of 30k (ablation) | Consist-DPO, Mulberry-2B · MMStar | accuracy: 49.27 -> 50.20 | Improved |
| apply ReCoT training | Mulberry-7B SFT · ScienceQA | CSR_self: 22.29 -> 75.00 | Improved |
Closest papers by meaning
- Language Bias in LVLMs: From In-Depth Analysis to Simple and Effective Mitigation (ICML 2026 · similarity 0.89)
- VISCO: Benchmarking Fine-Grained Critique and Correction Towards Self-Improvement in Visual Reasoning (CVPR 2025 · similarity 0.89)
- Towards Self-Refinement of Vision-Language Models with Triangular Consistency (NeurIPS 2025 · similarity 0.89)
- Sherlock: Self-Correcting Reasoning in Vision-Language Models (NeurIPS 2025 · similarity 0.89)
- Enhancing Large Vision Language Models with Self-Training on Image Comprehension (NeurIPS 2024 · similarity 0.88)
- DRESS: Instructing Large Vision-Language Models to Align and Interact with Humans via Natural Language Feedback (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
