SELF: Self-Evolution with Language Feedback

TL;DR

SELF is a novel framework that enables large language models to self-improve through iterative self-reflection, self-feedback, and self-refinement, leading to enhanced performance without human intervention.

Contribution

The paper introduces SELF, a new method allowing LLMs to autonomously self-evolve via self-reflection and iterative fine-tuning, advancing beyond traditional static models.

Findings

SELF improves LLM performance in mathematics tasks.

SELF enhances general task responses without human-labeled data.

Iterative self-refinement leads to progressive model improvement.

Abstract

Large Language Models (LLMs) have demonstrated remarkable versatility across various domains. To further advance LLMs, we propose 'SELF' (Self-Evolution with Language Feedback), a novel approach that enables LLMs to self-improve through self-reflection, akin to human learning processes. SELF initiates with a meta-skill learning process that equips the LLMs with capabilities for self-feedback and self-refinement. Subsequently, the model undergoes an iterative process of self-evolution. In each iteration, it utilizes an unlabeled dataset of instructions to generate initial responses. These responses are enhanced through self-feedback and self-refinement. The model is then fine-tuned using this enhanced data. The model undergoes progressive improvement through this iterative self-evolution process. Moreover, the SELF framework enables the model to apply self-refinement during inference, which further improves response quality. Our experiments in mathematics and general tasks demonstrate that SELF can enhance the capabilities of LLMs without human intervention. The SELF framework indicates a promising direction for the autonomous evolution of LLMs, transitioning them from passive information receivers to active participants in their development.