Self-JGAR: Self-Judgement For Generative Adversarial Reasoning in Large Language Models

Anonymous

Abstract: Large Language Models (LLMs) leverage their output for refinement, attracting increasing interest in such techniques. However, the illusion issue makes it challenging to guarantee the effectiveness of this refinement. Incorporating external feedback is pivotal for addressing the challenges in the refinement to ensure the reliability of the generated content. We introduce a framework, Self-JGAR, which utilizes adversarial learning to update the judgment capacity of LLMs and steer LLMs reasoning in the right direction. The framework endows LLMs with the capability to make judgments about their reasoning process, thereby enhancing their reasoning ability. Experiment results show that our framework outperforms the strong baseline on reasoning tasks. The codes will be released upon the acceptance of this paper.

Paper Type: short

Research Area: Generation

Contribution Types: NLP engineering experiment

Languages Studied: English