Large Language Models (LLMs) are able to improve their responses when instructed to do so, a capability known as self-correction. When instructions provide only a general and abstract goal without specific details about potential issues in the response, LLMs must rely on their internal knowledge to improve response quality, a process referred to as intrinsic self-correction. The empirical success of intrinsic self-correction is evident in various applications, but how and why it is effective remains unknown. Focusing on moral self-correction in LLMs, we reveal a key characteristic of intrinsic self-correction: performance convergence through multi-round interactions; and provide a mechanistic analysis of this convergence behavior. Based on our experimental results and analysis, we uncover the underlying mechanism of convergence: consistently injected self-correction instructions activate moral concepts that reduce model uncertainty, leading to converged performance as the activated moral concepts stabilize over successive rounds. This paper demonstrates the strong potential of moral self-correction by showing that it exhibits a desirable property of converged performance.

Large Language Models (LLMs) are capable of producing content that perpetuates stereotypes, discrimination, and toxicity. The recently proposed moral self-correction is a computationally efficient method for reducing harmful content in the responses of LLMs. However, the process of how injecting self-correction instructions can modify the behavior of LLMs remains under-explored. In this paper, we explore the effectiveness of moral self-correction by answering three research questions: (1) In what scenarios does moral self-correction work? (2) What are the internal mechanisms of LLMs, e.g., hidden states, that are influenced by moral self-correction instructions? (3) Is intrinsic moral self-correction actually superficial? We argue that self-correction can help LLMs find a shortcut to more morally correct output, rather than truly reducing the immorality stored in hidden states. Through empirical investigation with tasks of language generation and multi-choice question answering, we conclude: (i) LLMs exhibit good performance across both tasks, and self-correction instructions are particularly beneficial when the correct answer is already top-ranked; (ii) The morality levels in intermediate hidden states are strong indicators as to whether one instruction would be more effective than another; (iii) Based on our analysis of intermediate hidden states and task case studies of self-correction behaviors, we are first to propose the hypothesis that intrinsic moral self-correction is in fact superficial.

Large Language Models (LLMs) can improve their responses when instructed to do so, a capability known as self-correction. When these instructions lack specific details about the issues in the response, this is referred to as leveraging the intrinsic self-correction capability. The empirical success of self-correction can be found in various applications, e.g., text detoxification and social bias mitigation. However, leveraging this self-correction capability may not always be effective, as it has the potential to revise an initially correct response into an incorrect one. In this paper, we endeavor to understand how and why leveraging the self-correction capability is effective. We identify that appropriate instructions can guide LLMs to a convergence state, wherein additional self-correction steps do not yield further performance improvements. We empirically demonstrate that model uncertainty and activated latent concepts jointly characterize the effectiveness of self-correction. Furthermore, we provide a mathematical formulation indicating that the activated latent concept drives the convergence of the model uncertainty and self-correction performance. Our analysis can also be generalized to the self-correction behaviors observed in Vision-Language Models (VLMs). Moreover, we highlight that task-agnostic debiasing can benefit from our principle in terms of selecting effective fine-tuning samples. Such initial success demonstrates the potential extensibility for better instruction tuning and safety alignment.