Enhancing Moral Diagnosis and Correction in Large Language Models

TL;DR

This paper presents a pragmatic inference-based method to improve moral error detection and correction in large language models, demonstrating broad applicability and superior performance across diverse moral and social tasks.

Contribution

It introduces a unifying variable, pragmatic inference load, enabling generalization across tasks and enhancing LLMs' moral diagnostic and correction capabilities.

Findings

Outperforms baseline methods in moral error correction

Generalizes across diverse moral and social tasks

Improvements stem from learned inferential processes

Abstract

Identifying specific moral errors in an input and generating appropriate corrections require moral sensitivity in large language models (LLMs), which is fundamental for developing their moral performance, yet a challenging task. This study leverages a pragmatic inference-based approach to enhance both the moral diagnosis and corrections of models. Crucially, our method generalizes across a diverse set of different tasks, including moral reasoning, toxic language detection, social bias detection, and jailbreaks, despite substantial differences in their semantic formulations. To enable such generalization, the study also introduces a unifying variable, pragmatic inference load, which captures the degree of pragmatic reasoning required across tasks. Experimental results show that our approach enables LLMs to produce high-quality diagnostic outputs of moral errors, make effective corrections, and consistently outperform a range of baseline methods. Further analyses reveal that these improvements do not arise from heuristic-based response patterns, but from learned inferential processes, highlighting the effectiveness of our approach.