Aligning LLM Uncertainty with Human Disagreement in Subjectivity Analysis

2026-05-11

AI Summary

This work addresses the limitations of conventional subjectivity analysis models that rely on aggregated labels while ignoring inherent disagreements among human annotators, leading to overconfidence on low-consensus samples and undermining reliability and generalization. To remedy this, the authors propose DPUA, an uncertainty-aware framework that explicitly models human disagreement as an uncertainty signal. DPUA employs a two-stage process—disagreement perception and uncertainty alignment—integrating adaptive decoupled learning, GRPO-based reinforcement optimization, and multi-task joint training for label prediction, rationale generation, and uncertainty estimation. Experimental results demonstrate that DPUA maintains competitive task performance while significantly mitigating overconfidence, enhancing out-of-distribution generalization, and achieving strong alignment between model uncertainty and human disagreement distributions.

Abstract

Large language models for subjectivity analysis are typically trained with aggregated labels, which compress variations in human judgment into a single supervision signal. This paradigm overlooks the intrinsic uncertainty of low-agreement samples and often induces overconfident predictions, undermining reliability and generalization in complex subjective settings. In this work, we advocate uncertainty-aware subjectivity analysis, where models are expected to make predictions while expressing uncertainty that reflects human disagreement. To operationalize this perspective, we propose a two-phase Disagreement Perception and Uncertainty Alignment (DPUA) framework. Specifically, DPUA jointly models label prediction, rationale generation, and uncertainty expression under an uncertainty-aware setting. In the disagreement perception phase, adaptive decoupled learning enhances the model's sensitivity to disagreement-related cues while preserving task performance. In the uncertainty alignment phase, GRPO-based reward optimization further improves uncertainty-aware reasoning and aligns the model's confidence expression with the human disagreement distribution. Experiments on three subjectivity analysis tasks show that DPUA preserves task performance while better aligning model uncertainty with human disagreement, mitigating overconfidence on boundary samples, and improving out-of-distribution generalization.

Problem

• subjectivity analysis
• human disagreement
• model uncertainty
• overconfidence
• label aggregation

Innovation

• uncertainty alignment
• human disagreement
• subjectivity analysis
• adaptive decoupled learning
• GRPO-based reward optimization