Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2510.07290 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866918172087549952 |
|---|---|
| author | Liu, Guangliang Mao, Haitao Cao, Bochuan Xue, Zhiyu Zhang, Xitong Wang, Rongrong Johnson, Kristen Marie |
| author_facet | Liu, Guangliang Mao, Haitao Cao, Bochuan Xue, Zhiyu Zhang, Xitong Wang, Rongrong Johnson, Kristen Marie |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2510_07290 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | On the Convergence of Moral Self-Correction in Large Language Models Liu, Guangliang Mao, Haitao Cao, Bochuan Xue, Zhiyu Zhang, Xitong Wang, Rongrong Johnson, Kristen Marie Computation and Language Machine Learning 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. |
| title | On the Convergence of Moral Self-Correction in Large Language Models |
| topic | Computation and Language Machine Learning |
| url | https://arxiv.org/abs/2510.07290 |