💬 NLP

Aligning LLM Uncertainty with Human Disagreement in Subjectivity Analysis

This paper proposes the Disagreement Perception and Uncertainty Alignment (DPUA) framework, a two-phase approach that enhances large language models' subjectivity analysis by explicitly modeling human disagreement to align model uncertainty with real-world judgment variations, thereby reducing overconfidence and improving generalization.

Original authors: Junyu Lu, Deyi Ji, Xuanyi Liu, Lanyun Zhu, Bo Xu, Liang Yang, Xian-Sheng Hua, Hongfei Lin

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Junyu Lu, Deyi Ji, Xuanyi Liu, Lanyun Zhu, Bo Xu, Liang Yang, Xian-Sheng Hua, Hongfei Lin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Core Problem: The "Overconfident Expert"

Imagine you ask a super-smart robot (a Large Language Model, or LLM) to judge a joke.

The Robot's Old Way: The robot was trained by looking at thousands of examples where humans voted on whether something was "funny" or "not funny." If 90% of humans said "funny," the robot learned that it is 100% funny. If 51% said "funny," the robot still learned it is "funny," but it acts like it's 100% sure.
The Flaw: In subjective tasks (like detecting sarcasm, offense, or sentiment), humans often disagree. One person might find a comment offensive, while another finds it funny. When we force the robot to pick just one answer, it ignores the fact that humans are confused. The result? The robot becomes overconfident. It gives a definitive answer on a tricky question, even when the question itself is ambiguous. It's like a weather forecaster saying "It will definitely rain" when the sky is actually a mix of clouds and sun.

The Solution: DPUA (The "Humble Detective")

The authors propose a new framework called DPUA (Disagreement Perception and Uncertainty Alignment). Instead of just teaching the robot what the answer is, they teach it how unsure it should be when humans disagree.

Think of DPUA as training a detective in two specific phases:

Phase 1: Learning to Spot the "Fog" (Disagreement Perception)

In this phase, the robot learns to recognize when a situation is "foggy" (ambiguous).

The Analogy: Imagine a teacher grading essays. For clear-cut essays (high agreement), the teacher gives a straight A. But for essays where the student's intent is unclear (low agreement), the teacher doesn't just give a grade; they write a note saying, "This is tricky. Some people might see it as X, others as Y."
How it works: The robot is trained to pay extra attention to these "foggy" cases. It learns to generate a "rationale" (an explanation) that admits, "I see why some people think this is offensive, but I also see why others might not." It stops trying to force a single, rigid answer on messy data.

Phase 2: Tuning the "Confidence Dial" (Uncertainty Alignment)

In this phase, the robot learns to adjust its "confidence dial" to match how much humans actually disagree.

The Analogy: Imagine a scale that measures how sure you are.
- If 100 people agree a statement is true, the robot turns the dial to 100%.
- If only 51 people agree (and 49 disagree), the robot turns the dial down to 55%.
- If the humans are split 50/50, the robot turns the dial to 50% (the lowest it goes), effectively saying, "I have no idea, and that's okay."
How it works: The researchers use a special reward system (like a video game score) to punish the robot if it acts too confident when humans are confused, and reward it if it acts cautious when the data is messy.

What Did They Find?

The paper tested this on three tasks: detecting sarcasm, spotting offensive language, and analyzing sentiment (positive/negative).

It didn't lose its smarts: The robot didn't get worse at giving the right answer. In fact, it stayed just as good at predicting the label.
It became more honest: The robot stopped pretending to be sure when it shouldn't be. On tricky "boundary" cases (where humans disagree), the robot's confidence score dropped to match the human confusion.
It generalized better: When the robot was tested on new types of subjective problems it hadn't seen before, it handled them better than the old models. It learned a general skill of "knowing what it doesn't know."

The Takeaway

The paper argues that for subjective tasks, uncertainty is not a bug; it's a feature.

By teaching AI to express uncertainty in a way that mirrors human disagreement, we get models that are more reliable. Instead of a robot that confidently gives the wrong answer on a tricky joke, we get a robot that says, "This is likely a joke, but I can see why some people might be offended, so I'm only 60% sure." This makes the AI a better partner for humans in complex, nuanced decision-making.

Technical Summary: Aligning LLM Uncertainty with Human Disagreement in Subjectivity Analysis

Problem Statement
Current Large Language Models (LLMs) for subjectivity analysis (e.g., sentiment analysis, sarcasm detection, offensiveness detection) are typically trained on aggregated labels derived from majority voting. This paradigm compresses the intrinsic variations in human judgment into a single supervision signal. Consequently, it overlooks the uncertainty inherent in low-agreement samples, often inducing overconfident predictions and rigid decision boundaries. This limitation undermines model reliability and generalization in complex subjective settings where semantic boundaries are fuzzy and human interpretations vary. The core problem addressed is how to enable models to not only predict labels but also express uncertainty that accurately reflects the distribution of human disagreement.

Methodology: The DPUA Framework
The authors propose a two-phase framework called Disagreement Perception and Uncertainty Alignment (DPUA) to operationalize uncertainty-aware subjectivity analysis. The framework requires models to jointly perform label prediction, rationale generation, and confidence estimation.

Phase 1: Disagreement Perception (Adaptive Decoupled Learning)
- Goal: Enhance the model's sensitivity to disagreement signals while preserving task performance.
- Mechanism: The authors introduce an adaptive decoupled learning strategy during Supervised Fine-Tuning (SFT). The training objective is reweighted based on the human agreement score ( $C^h_i$ $C_{i}^{h}$ ) of each sample.
  - High-agreement samples: Receive stronger supervision for label prediction to ensure stable decision patterns.
  - Low-agreement samples: Receive increased weight for rationale generation to encourage the model to learn disagreement-related cues and evidence.
- Output: The model learns to generate rationales that identify both label justifications and potential sources of ambiguity or disagreement.
Phase 2: Uncertainty Alignment (GRPO-based Optimization)
- Goal: Refine uncertainty-aware reasoning and align the model's confidence expression with the human disagreement distribution.
- Mechanism: The authors employ Group Relative Policy Optimization (GRPO). This phase optimizes the model using two specific reward components:
  - Uncertainty-aware Reasoning Reward: Evaluates the quality of the generated rationale (label justification and disagreement-cue identification) using a judge model (GPT-4o-mini). The weighting of these components is dynamically adjusted based on the human agreement score.
  - Uncertainty Calibration Reward: Measures the discrepancy between the model's predicted confidence distribution and the empirical human disagreement distribution using Mean Absolute Error (MAE). This reward encourages the model to be less confident on low-agreement samples and more confident on high-agreement ones.
- Optimization: The model is updated by maximizing the group-relative advantage of sampled outputs, fostering reasoning that aligns with human uncertainty patterns.

Key Contributions

Paradigm Shift: The paper advocates for "uncertainty-aware subjectivity analysis," moving beyond fitting a single aggregated label to a setting where models must identify disagreement cues and express calibrated uncertainty.
Framework Proposal: The introduction of DPUA, a two-phase framework that jointly models label prediction, rationale generation, and confidence estimation.
Empirical Validation: Systematic evaluation on three subjectivity analysis tasks (sarcasm detection, sentiment analysis, and offensiveness detection) demonstrating that DPUA preserves core task performance while significantly improving the alignment between model uncertainty and human disagreement.

Experimental Results
Experiments were conducted on three datasets: Conversational Sarcasm Corpus (CSC), GoEmotions (GOE), and MD-Agreement (MD), using Qwen-3 and LLaMa-3.1 backbones.

Uncertainty Alignment: DPUA significantly outperformed baselines (zero-shot, few-shot, standard SFT, and vanilla GRPO) in uncertainty alignment metrics. It reduced the Mean Absolute Error (MAE) between model and human distributions by approximately 5.94% to 8.31% and improved Pearson correlation coefficients by 5.13% to 6.06% across backbones.
Task Performance: Unlike standard uncertainty-aware baselines that sometimes sacrificed accuracy, DPUA maintained competitive label prediction performance, showing slight improvements in Accuracy and F1 scores over strong baselines.
Overconfidence Mitigation: Visualization of confidence distributions showed that DPUA mitigates overconfidence on boundary samples, assigning intermediate probability values to low-agreement cases where standard SFT tends to be overly certain.
Generalization: In out-of-distribution (OOD) transfer settings (e.g., training on GOE and testing on CSC), DPUA demonstrated superior generalization, achieving the best or competitive results in both label prediction and uncertainty alignment, suggesting the learned uncertainty-aware patterns are transferable.
Explainability: Case studies and rationale evaluations indicated that DPUA generates more informative and sound rationales that explicitly reflect ambiguity and potential human disagreement.

Significance and Claims
The paper claims that modeling human disagreement as an intrinsic source of uncertainty is crucial for developing trustworthy LLMs in subjective decision-making scenarios. By aligning model confidence with the actual distribution of human disagreement, DPUA addresses the issue of overconfidence on boundary samples, which is a common failure mode in current subjectivity analysis systems. The authors assert that their approach provides a more reliable foundation for subjective tasks by ensuring that models express uncertainty in a manner consistent with the complexity of human cognition, rather than forcing a false sense of certainty through aggregated labels. The work highlights that uncertainty alignment is a distinct challenge from label prediction and requires specific architectural and training strategies to achieve effectively.