In standard supervised learning, we treat human annotations as a stepping stone to a single "ground truth." If five radiologists look at a medical scan, or five crowd-workers rate the sentiment of a tweet, standard practice dictates collapsing their opinions into a single hard label via majority voting or averaging.
This aggregation step relies on a fragile assumption: that variance among human annotators is merely noise to be filtered out. In reality, human disagreement is often a vital signal. When we force a consensus, we discard rich information about task difficulty, linguistic ambiguity, and subjective interpretation. A growing body of research demonstrates that training models directly on the full distribution of human annotations—rather than an aggregated consensus—fundamentally changes what and how a model learns.
The Information Loss of Majority Voting
To understand why label aggregation hurts model performance, consider how a standard classification model is trained using cross-entropy loss against a one-hot target vector.
Imagine an image of a borderline medical case where three out of five experts diagnose a condition as "malignant" and two diagnose it as "benign." Under a majority-vote paradigm, the target becomes a hard one-hot vector: [1.0, 0.0]. The loss function penalizes the model unless it predicts the malignant class with 100% confidence. As a result, the model is forced to learn absolute certainty for an inherently ambiguous case.
As Ben Glocker et al. observe in Spatially Varying Label Smoothing: Capturing Uncertainty from Expert Annotations, this reductionist approach strips away the spatial and contextual nuances of where and why experts disagree. By collapsing these diverse perspectives, models become systematically overconfident, failing to capture the aleatoric uncertainty embedded in the data.
This problem is not unique to computer vision. In Natural Language Inference (NLI), sentences are frequently open to multiple valid interpretations. To highlight this, Yixin Nie et al. introduced the ChaosNLI dataset in What Can We Learn from Collective Human Opinions on Natural Language Inference Data?. By collecting 100 annotations per example, the authors revealed that state-of-the-art models trained on majority-vote labels perform no better than random guess on examples where human consensus is divided. The models simply lack the capacity to recognize when an inference is genuinely ambiguous.
Standard Pipeline:
[Annotator 1: A]
[Annotator 2: A] --> Majority Vote --> Hard Label [A] --> Information Lost
[Annotator 3: B]
Distributional Pipeline:
[Annotator 1: A]
[Annotator 2: A] --> Label Distribution --> Soft Target [0.66 A, 0.33 B] --> Information Preserved
[Annotator 3: B]
Why Soft Labels Build Better Representations
When we preserve the full distribution of human opinions as a "soft label" (e.g., target probabilities of [0.6, 0.4] instead of [1.0, 0.0]), the learning dynamics shift in three fundamental ways:
1. Regularization and Calibration
Soft labels act as data-driven regularizers. Instead of pushing network weights to output extreme logits, soft targets smooth the optimization landscape. In A Case for Soft Loss Functions, Alexandra Uma et al. demonstrate that training on the full distribution of crowd judgments significantly improves both model performance and uncertainty calibration across natural language tasks. Because the model is penalized for being overconfident on disputed items, its predicted probabilities naturally align with actual human uncertainty.
2. Geometry of the Feature Space
Forcing a model to separate ambiguous items into strict binary categories distorts its internal representation space. By utilizing soft labels, models can place ambiguous or borderline examples in the geometric transitions between class clusters. Biqiao Zhang et al. leverage this in $f$-Similarity Preservation Loss for Soft Labels: A Demonstration on Cross-Corpus Speech Emotion Recognition, showing that forcing deep metric learning frameworks to preserve the similarity matrix of soft labels yields representations that generalize far better to entirely unseen datasets.
3. Disagreement as Privileged Information
Human disagreement can serve as an explicit signal of instance difficulty. In Ambiguity Helps: Classification with Disagreements in Crowdsourced Annotations, Viktoriia Sharmanska et al. treat the variance in crowd annotations as "privileged information" available during training. This variance guides the model to identify which features make an example easy or hard, ultimately improving test-time accuracy even when annotator distributions are unavailable.
Architectural Strategies for Disagreement
Researchers have developed several distinct paradigms to bypass the information loss of label aggregation:
Soft Loss Functions
The most direct approach is modifying the loss function to accept continuous distributions. For instance, Zifu Wang et al. developed Dice Semimetric Losses: Optimizing the Dice Score with Soft Labels to allow medical image segmentation models to optimize directly for multi-rater distributions, bypassing the traditional requirement for binary targets. Similarly, Alberto Purpura introduces probabilistic loss functions in Learning to Rank from Relevance Judgments Distributions to capture the full variance of human relevance judgments in information retrieval.
Multi-Task Learning
Rather than replacing the hard label, some frameworks predict both the consensus and the disagreement distribution simultaneously. In Beyond Black & White: Leveraging Annotator Disagreement via Soft-Label Multi-Task Learning, Tommaso Fornaciari et al. design a multi-task network where one head predicts the hard majority vote and an auxiliary head predicts the full annotator distribution. This architecture forces the shared layers to learn features that explain why humans disagree, preventing the model from overfitting to noisy or narrow majorities.
Modeling Individual Annotators
Instead of merging labels beforehand, models can learn the individual biases of each annotator. Melody Y. Guan et al. pioneered this in Who Said What: Modeling Individual Labelers Improves Classification by appending a rater-specific layer to a deep neural network. This allows the base model to learn from the raw, unaggregated annotations while the rater layer learns to model each expert's unique reliability and bias.
A similar mechanism is used by Yichong Leng et al. in MBNet: MOS Prediction for Synthesized Speech with Mean-Bias Network, which decouples speech quality prediction into a mean score network and a rater bias network, maximizing the training signal extracted from limited human evaluations.
Application Domains
The shift from hard consensus to soft distributions has proven critical in fields where data is inherently subjective or high-stakes.
- Medical Diagnosis: In tasks like tumor segmentation or cancer grading, inter-observer variability is immense. Gesa Mittmann et al. demonstrate in Pathologist-like explainable AI for interpretable Gleason grading in prostate cancer that training on soft labels from dozens of international pathologists allows models to preserve borderline clinical cases, mirroring human reasoning and improving diagnostic safety.
- Subjective NLP: Tasks like hate speech detection, irony detection, and sentiment analysis rarely have a single objective truth. As Paul Röttger et al. argue in Two Contrasting Data Annotation Paradigms for Subjective NLP Tasks, preserving the variance in human labels via descriptive annotation provides a more accurate reflection of pluralistic societal perceptions.
- Evaluating LLMs: Automated evaluation metrics often fail because they assume a single correct answer. Kai Ma demonstrates in Is LLM an Overconfident Judge? Unveiling the Capabilities of LLMs in Detecting Offensive Language with Annotation Disagreement that large language models trained on aggregated data suffer from severe overconfidence, but exposing them to disagreement distributions aligns them far better with human-like nuance.
Conclusion
The transition toward learning from human disagreement represents a fundamental shift in data-centric AI. By treating annotator variance as a rich, probabilistic signal rather than noise, we move away from brittle, overconfident models. Preserving the distribution of human opinions allows machine learning systems to embrace the gray areas of human language and perception, resulting in models that are better calibrated, more robust to distribution shifts, and inherently more transparent.