Self-correction for OOD generalization

Vanya Bannihatti Kumar, Abhinav Sukumar Rao, Aditi Raghunathan

Keywords: Self-correction; OOD generalization

TL;DR: We use reasoning based methods for OOD generalization in LLMs and VLMs

Abstract:

In this work, we aim to study how the self-correction mechanisms aid OOD (out-of-distribution) generalization in both multimodal and language-only models. Reasoning based methods like self-refine and STaR have helped to improve the correction capacity of the language models; however there have been no studies quantifying the reasoning improvement to help OOD generalization of these models.
Initial results, show an improvement of 1.6%-2% on an OOD dataset where the model is finetuned using either self-refinement or STaR on an ID (in-distribution) dataset.

Submission Number: 68