Human Label Variation as Stable Signal: Learning Annotator-Specific Explanation Behavior via Cross-Annotator Preference Optimization

2026-05-27 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

The study "Human Label Variation as Stable Signal" investigates whether large language models (LLMs) can learn and replicate annotator-specific reasoning and preferences from free-text explanations. Researchers analyzed human label variation (HLV) across two sentence-pair tasks—natural language inference and paraphrase judgment—each involving four annotators. They observed that individual annotator patterns, while weak at the single-annotation level due to strong input-content effects, become discernible after input-content reduction and aggregation. The work compared prompting and supervised fine-tuning (SFT) baselines, then introduced Cross-Annotator Preference Optimization (CAPO). Experiments demonstrated that prompting is unstable, SFT improves behavior capture, and CAPO further enhances aggregation-aware imitation and judge-based attribution, maintaining target-specific reasoning patterns. This suggests HLV can be a stable signal for scalable, explanation-based annotation.

Key takeaway

For machine learning engineers developing explanation-based annotation systems, this research indicates that utilizing annotator-specific reasoning patterns can significantly improve model fidelity. Your teams should consider moving beyond simple label agreement by incorporating methods like Cross-Annotator Preference Optimization (CAPO) to train LLMs on individual annotator histories. This approach offers a path to more scalable and nuanced data labeling, potentially reducing annotation costs and improving explanation quality.

Key insights

Large language models can learn and reproduce individual annotator explanation styles from human label variation.

Principles

• HLV reveals annotator reasoning.
• Annotator patterns stabilize with aggregation.
• Preference optimization enhances imitation.

Method

Cross-Annotator Preference Optimization (CAPO) contrasts a target annotator's response with other valid but less target-specific annotations for the same input to learn individual explanation behavior.

In practice

• Use SFT for initial behavior capture.
• Apply CAPO for refined imitation.
• Ground annotations in annotator histories.

Topics

• Human Label Variation
• Large Language Models
• Cross-Annotator Preference Optimization
• Supervised Fine-Tuning
• Natural Language Inference
• Explanation-based Annotation

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Quantifying and Predicting Disagreement in Graded Human Ratings — Annotation disagreement in graded human ratings can be predicted from textual features, especially for inappropriate language.
• Learning to Label: A Reinforced Self-Evolving Framework for Semi-supervised Referring Expression Segmentation — A reinforced self-evolving framework improves semi-supervised referring expression segmentation by learning to construct reliable pseudo-labels.
• Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability — Reasoning SFT can generalize cross-domain, but it is conditional on optimization, data quality, and model capability.
• Selective Prediction from Agreement: A Lipschitz-Consistent Version Space Approach — Selective prediction can be achieved by identifying labels forced by agreement among Lipschitz-consistent classification heads.
• Metric-Dependent Annotation Saturation for Learning from Label Distributions — Annotation budget needs depend on the target evaluation metric, as different metrics saturate at varying annotator counts.
• A Causal Argumentation Method for Explainability of Machine Learning Models — Integrating causal discovery with argumentation frameworks provides "why" explanations for ML model predictions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.