An Assessment of Human vs. Model Uncertainty in Soft-Label Learning and Calibration

The gist: Human soft-labels improve model calibration and robustness by acting as a regularizer that aligns model uncertainty with human uncertainty, which synthetic labels fail to do.

How it works

Human-elicited soft-labels encode the diverse origins of uncertainty that characterize real-world data and have been shown to directly improve both model calibration and robustness<ref:2605.18648#pg2>. Human uncertainty is multifaceted, originating from varied interpretations of annotation schemes, differences in cultural or background knowledge, cognitive difficulties in separating similar concepts, or data-specific issues like image occlusion and noise<ref:2605.18648#pg2>. Prior studies demonstrating accuracy gains from human softlabels conflate the soft-supervision with the implicit correction of the underlying label mode (i.e. shifting the argmax to a more accurate class during re-annotation)<ref:2605.18648#pg3>. Decoupling these effects requires a controlled testbed, akin to the use of MNIST to validate hypotheses in generalization or challenge assumptions regarding adversarial robustness<ref:2605.18648#pg3>.

Dataset and Annotation Scheme

The researchers conducted experiments on two digit datasets (MNIST and a synthetic equivalent), re-annotating subsets to explicitly capture individual human class uncertainty<ref:2605.18648#pg3>. They evaluated four digit-based datasets: the standard MNIST, the Swedish historical handwritten dataset ARDIS, and two synthetic variants of Ambiguous-MNIST<ref:2605.18648#pg4>. Difficulty regions are defined by thresholds on confidence (µ) and variability (σ)<ref:2605.18648#pg16>. To capture human uncertainty without forcing arbitrary numerical guesses, they adopted a forced-choice yes/no approach augmented with an explicit unsure option<ref:2605.18648#pg4>.

Evaluation Metrics and Findings

The evaluation focuses on two main aspects: Test-set performance and Training dynamics<ref:2605.18648#pg4>. They use Kullback-Leibler Divergence (KLD) as the main diagnostic to measure the discrepancy between human-annotated soft-labels and model predictions, measuring DKL(Phuman∥Pmodel)<ref:2605.18648#pg4>. They also report the Brier Score<ref:2605.18648#pg4>. On the Mukhoti dataset, models trained on synthetic (orig.) labels show substantially higher accuracy on the HLV subset than human-trained counterparts, but this trend inverts when evaluated against new human labels<ref:2605.18648#pg18>.

Key Results and Alignment

Models supervised by soft-labels achieve significantly lower KLD on HLV subsets compared to those trained on orig./synth. or maj.n targets<ref:2605.18648#pg4>. Training under sof tw shows the model is learning to be uncertain when humans are uncertain, which aligns with human uncertainty<ref:2605.18648#pg10>. The Spearman correlation between late stage model training dynamics and human uncertainty (uprop) is strongest for sof tw, showing meaningful negative correlations with model confidence<ref:2605.18648#pg10>. This suggests that models trained on soft-labels mirror human uncertainty throughout the learning process<ref:2605.18648#pg8>.

Conclusion and Limitations

The work provides a diagnostic testbed for human-AI uncertainty alignment, demonstrating that while human soft-labels do provide accuracy gains, their larger value lies in acting as a regularizer that improves model calibration on difficult samples and promotes stable convergence across training runs<ref:2605.18648#pg3>. The findings suggest that synthetic proxies fail to encode human uncertainty, reinforcing the specialized role of human-grounded evaluation<ref:2605.18648#pg10>. The study is limited by restricting the testbed to simple digits and smaller models to isolate perceptual ambiguity, and results should be treated strictly as a research testbed rather than a production-grade training set<ref:2605.18648#pg10>.

--- Page 1 ---

An Assessment of Human vs. Model Uncertainty in Soft-Label Learning and Calibration Maja Pavlovic Queen Mary University London m.pavlovic@qmul.ac.uk Silviu Paun spaun3691@gmail.com Massimo Poesio Queen Mary University London - University of Utrecht m.poesio@qmul.ac.uk,uu.nl Abstract Central to human-aligned AI is understanding the benefits of human-elicited labels over synthetic alternatives<ref:2605.18648#pg2>. While human soft-labels improve calibration by capturing uncertainty, prior studies conflate these benefits with the implicit correction of mislabeled data (mode shifts), obscuring true effects of soft-labels<ref:2605.18648#pg3>. We present a controlled audit of soft-label learning across MNIST and a synthetic variant, re-annotating subsets to extract human uncertainty<ref:2605.18648#pg3>. By decoupling soft-label supervision from underlying label mode shifts, we show that while human soft-labels do provide accuracy gains, their larger value lies in acting as a regularizer that improves model calibration on difficult samples and promotes stable convergence across training runs<ref:2605.18648#pg3>. Dataset cartography reveals models trained on human softlabels mirror human uncertainty, whereas those trained on synthetic labels fail to align with humans<ref:2605.18648#pg3>. Broadly, this work provides a diagnostic testbed for human-AI uncertainty alignment<ref:2605.18648#pg3>. 1 Introduction Model calibration ensures that predicted probabilities accurately reflect the true likelihood of outcomes<ref:2605.18648#pg2>. Traditionally, models are trained and evaluated using a one-hot encoded vectors<ref:2605.18648#pg2>; however, as model complexity increases, models often become less calibrated<ref:2605.18648#pg2>. While early approaches demonstrate that soft-labels enhance calibration and generalization<ref:2605.18648#pg2>, recent work suggests that synthetically generated data can further refine calibration by exposing models to a wider variety of edge cases<ref:2605.18648#pg2>. However, using teacher model soft-labels or otherwise algorithmically generated soft-labels risks propagating synthetic biases rather than reflecting the nuances of human uncertainty<ref:2605.18648#pg2>. In contrast to these synthetic soft-labels, human-elicited soft-labels encode the diverse origins of uncertainty that characterize real-world data and have been shown to directly improve both model calibration and robustness<ref:2605.18648#pg2>. Human uncertainty is multifaceted, originating from varied interpretations of annotation schemes, differences in cultural or background knowledge, cognitive difficulties in separating similar concepts, or data-specific issues like image occlusion and noise<ref:2605.18648#pg2>. Additionally, difficulty arises when instances legitimately belong to multiple categories or are inherently vague<ref:2605.18648#pg2>.

--- Page 2 ---

due to ambiguity rather than overlapping categories [25, 40]. Recent efforts have begun integrating individual human uncertainty into annotated datasets<ref:2605.18648#pg4>. While the utility of soft-labels is thought to depend on their alignment with human label variation<ref:2605.18648#pg4>, our fundamental understanding of how and why human-derived labels influence model training dynamics remains limited<ref:2605.18648#pg4>. Prior studies demonstrating accuracy gains from human softlabels conflate the soft-supervision with the implicit correction of the underlying label mode (i.e. shifting the argmax to a more accurate class during re-annotation)<ref:2605.18648#pg4>. Decoupling these effects requires a controlled testbed, akin to the use of MNIST to validate hypotheses in generalization or challenge assumptions regarding adversarial robustness<ref:2605.18648#pg4>. To provide a similarly principled investigation into human soft-labels, we conduct experiments on two digit datasets (MNIST and a synthetic equivalent), re-annotating subsets to explicitly capture individual human class uncertainty<ref:2605.18648#pg4>. Within these testbeds, we restrict model capacity across three tractable baselines to capture fundamental learning differences<ref:2605.18648#pg4>.

1. A model-based system can be improved by incorporating human uncertainty signals to enhance calibration on difficult samples and promote stable convergence across training runs, as demonstrated by "we show that while human soft-labels do provide accuracy gains, their larger value lies in acting as a regularizer that improves model calibration on difficult samples and promotes stable convergence across training runs."

2. The improved system will exhibit better predictive uncertainty alignment by mirroring human uncertainty throughout the learning process, leading to models whose internal representations are more aligned with human-centric uncertainty than those trained on standard one-hot labels, and notably more than those trained on synthetic labels.

3. By utilizing uncertainty-weighted (sof tw) soft-labels instead of equally-weighted ones, the system will be better at capturing class likelihoods, as this scheme reflects a lower class-likelihood for Unsure selections compared to the equally-weighted scheme.

4. The system's training dynamics can be stabilized by leveraging soft-label supervision, as models trained under sof tw show a smoother optimization landscape, without the phase transitions visible in the majn supervision (esp. on the HLV subset).

5. The model will better generalize to ambiguous cases by being trained on human soft-labels, which are shown to have a substantial share of the accuracy gains on MNIST can be attributed to a shift in the label mode during re-annotation.

6. The system will exhibit superior calibration performance, as models supervised by soft-labels achieve significantly lower KLD on HLV subsets compared to those trained on orig./synth. or majn targets, which helps capture fine-grained information that is unrecoverable from collapsed one-hot representations.

7. The system can be designed to better handle label misalignment by prioritizing auditing difficult samples, as the paper notes that we show that ∼3% of accuracy gains on our MNIST subset stem from fixing label misalignment rather than architectural improvements.

Central to human-aligned AI is understanding the benefits of human-elicited labels over synthetic alternatives. While human soft-labels improve calibration by capturing uncertainty, prior studies conflate these benefits with the implicit correction of mislabeled data (mode shifts), obscuring true effects of soft-labels. We present a controlled audit of soft-label learning across MNIST and a synthetic variant, re-annotating subsets to extract human uncertainty. By decoupling soft-label supervision from underlying label mode shifts, we show that while human soft-labels do provide accuracy gains, their larger value lies in acting as a regularizer that improves model calibration on difficult samples and promotes stable convergence across training runs. Dataset cartography reveals models trained on human soft-labels mirror human uncertainty, whereas those trained on synthetic labels fail to align with humans. Broadly, this work provides a diagnostic testbed for human-AI uncertainty alignment.