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

The evolution of NLP has increasingly emphasized interpretability and human-centered AI. Early efforts focused on probabilistic label models and uncertainty quantification, exemplified by works like Nie et al. (2020, 2026) and Aroyo and Welty (2019). These approaches treated disagreement as noise, aiming to produce a consensus label distribution. However, recent studies highlight that disagreement often encodes meaningful differences in interpretation, reasoning, and perspective, especially in ambiguous tasks like NLI and paraphrase detection. The advent of explainable AI (XAI) and rationales has further enriched this understanding, with works such as Jiang and de Marneffe (2022) demonstrating the value of free-text explanations. Despite these advances, most models still rely on aggregate annotations, neglecting individual behavioral nuances. This gap motivates the current research, which seeks to explicitly model and learn individual explanation styles grounded in human annotation histories.

The core challenge addressed in this paper is the difficulty in capturing and reproducing stable, individual explanation behaviors from human annotations. Existing models often conflate content-driven explanations with stylistic tendencies, making it hard to disentangle the two. Single annotations are content-dominated and highly variable, which hampers the ability to learn consistent behavioral patterns. Moreover, current approaches lack mechanisms to leverage the natural variation among annotators as a source of meaningful signals. This results in models that produce generic or averaged explanations, limiting their interpretability and personalization. The problem is compounded by the scarcity of annotated data that explicitly targets individual behaviors, necessitating new methods to extract and utilize such signals effectively.

This paper introduces several key innovations:

• �� Content reduction via residual embeddings (E4) that diminish input influence, isolating stylistic features.
• �� Cross-annotator preference contrast: leveraging pairwise differences in responses from multiple annotators on the same input as supervision signals.
• �� The CAPO algorithm, which combines content reduction with preference contrastive learning, enabling models to learn stable, person-specific explanation behaviors.
• �� A comprehensive evaluation framework that includes automatic metrics (decision accuracy, explanation similarity, recognizability classifiers) and human validation, ensuring behavioral authenticity.
• �� Demonstrating that stable individual explanation behaviors can be learned and reproduced, opening new avenues for personalized NLP applications.

• �� Data Collection: Utilized two datasets—VariErr (NLI) and R2 (paraphrase)—each with four annotators providing labels and explanations.
• �� Data Preprocessing: Applied content reduction techniques, such as residual embeddings (E4), to minimize input content influence and highlight stylistic features.
• �� Baseline Methods: Implemented prompting strategies (in-context learning, profile prompting) and supervised fine-tuning (SFT) with independent adapters.
• �� CAPO Design:
• �� Input: Sentence pairs and target annotator ID.
• �� Pair Construction: For each input, select the target annotator’s response as positive; other annotators’ responses as negatives, forming preference pairs.
• �� Loss Function: Used a contrastive loss inspired by DPO (Rafailov et al.) to optimize the model’s ability to prefer target responses.
• �� Training: Initialized from SFT models, then fine-tuned with preference contrast, emphasizing target-specific behaviors.
• �� Evaluation: Employed decision accuracy, ROUGE/BERTScore for explanations, feature KL divergence, and a group classifier (GC) to assess recognizability; human validation was conducted for qualitative assessment.

The experiments involved training models on the VariErr and R2 datasets, with 300/100/100 splits for training, validation, and testing. Baselines included prompting (ICL, VP, VP-ICL), SFT, and the proposed CAPO. The models were evaluated on multiple metrics: label accuracy, explanation overlap (ROUGE-L, BERTScore), feature KL divergence, and recognizability scores from a group classifier. The analysis focused on the impact of group size (m) for aggregation, demonstrating that larger groups better reveal stable behaviors. Human evaluators also assessed explanation quality and style consistency. Results consistently showed CAPO’s superiority in modeling individual behaviors while maintaining decision accuracy.

CAPO achieved over 96% accuracy in recognizing target annotator behaviors at the group level, far surpassing SFT (~55%) and prompting (~40%). Feature KL divergence decreased significantly, indicating reduced content bias and enhanced stylistic fidelity. Human validation confirmed explanations generated by CAPO aligned more closely with target annotator reasoning, with an 82.8% agreement rate. The model effectively balanced behavior imitation with decision performance, demonstrating that stable, individual explanation styles can be learned and reproduced reliably. These findings validate the core hypothesis that annotator-specific behaviors are learnable signals, especially when aggregated across multiple instances.

This methodology can be directly applied to improve the quality and interpretability of large-scale annotation systems, enabling personalized explanations in AI assistants, medical diagnostics, and legal reasoning. It facilitates the creation of models that adapt to individual reasoning styles, making AI outputs more trustworthy and user-aligned. Long-term, the approach could underpin adaptive learning environments, where AI systems evolve with user feedback, and support cross-cultural or multilingual annotation tasks by capturing diverse explanatory norms. It also offers a pathway toward more transparent AI, where understanding individual reasoning enhances user trust and system accountability.

The reliance on multiple annotations per input limits applicability in low-resource or real-time scenarios. The pairwise preference training increases computational overhead and complexity, requiring careful pair selection to avoid confounding label or score differences. The approach’s effectiveness diminishes when annotator behaviors are highly inconsistent or context-dependent. Additionally, the current focus on English datasets necessitates validation in multilingual settings. Future work should address these limitations by developing more efficient training strategies, exploring few-shot or unsupervised variants, and testing across diverse languages and domains.