Self-Choose: Leveraging Diverse Reasoning Solutions to Self-Correct Multimodal Large Language Models
Keywords: Multimodal Large Language Models, Self-Correct, Reasoning, Prompting
Abstract: In the past few years, Multimodal Large Language Models (MLLMs) have achieved remarkable advancements in reasoning while still suffering from mistakes. Some existing approaches on LLMs self-correct the answers without external feedback, proven limited in reasoning. We revisit these previous approaches and propose an improved effective strategy dubbed Self-Choose to teach MLLMs to utilize diverse reasoning solutions to self-correct reasoning. Our approach first employs various reasoning methods to generate candidate answers. Then, it evaluates them by comparing the reasoning processes and candidate answers to choose the optimal solution. Finally, it outputs the best candidate or reflects to generate an improved solution if all the answers are deemed inaccurate. We evaluate our method on multiple datasets with mainstream foundation models including LLaVA and Gemini. The extensive experiments show that Self-Choose achieves consistent improvements on different benchmarks and metrics. We hope this study will promote future research on self-correction and its application across various tasks.
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13802
Loading