# Metric-Dependent Annotation Saturation for Learning from Label Distributions

arXiv: https://arxiv.org/abs/2605.29797 · HTML: https://mlsanity.com/paper/2605.29797/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 18 of its 35 references that are in the map.
- First posted on arXiv 28 May 2026.

## Builds on (most cited first, 18 of 18)

- [Rethinking the Inception Architecture for Computer Vision](https://mlsanity.com/paper/1512.00567/) (2015)
- [On Calibration of Modern Neural Networks](https://mlsanity.com/paper/1706.04599/) (2017)
- [A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference](https://mlsanity.com/paper/1704.05426/) (2017)
- [A large annotated corpus for learning natural language inference](https://mlsanity.com/paper/1508.05326/) (2015)
- [DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing](https://mlsanity.com/paper/2111.09543/) (2021)
- [When Does Label Smoothing Help?](https://mlsanity.com/paper/1906.02629/) (2019)
- [Calibration of Pre-trained Transformers](https://mlsanity.com/paper/2003.07892/) (2020)
- [Human uncertainty makes classification more robust](https://mlsanity.com/paper/1908.07086/) (2019)
- [What Can We Learn from Collective Human Opinions on Natural Language Inference Data?](https://mlsanity.com/paper/2010.03532/) (2020)
- [DICES Dataset: Diversity in Conversational AI Evaluation for Safety](https://mlsanity.com/paper/2306.11247/) (2023)
- [Eliciting and Learning with Soft Labels from Every Annotator](https://mlsanity.com/paper/2207.00810/) (2022)
- [Don't Waste a Single Annotation: Improving Single-Label Classifiers Through Soft Labels](https://mlsanity.com/paper/2311.05265/) (2023)
- [Crowd-Calibrator: Can Annotator Disagreement Inform Calibration in Subjective Tasks?](https://mlsanity.com/paper/2408.14141/) (2024)
- [Towards Understanding Why Label Smoothing Degrades Selective Classification and How to Fix It](https://mlsanity.com/paper/2403.14715/) (2024)
- [Training and Evaluating with Human Label Variation: An Empirical Study](https://mlsanity.com/paper/2502.01891/) (2025)
- [Subjective Crowd Disagreements for Subjective Data: Uncovering Meaningful CrowdOpinion with Population-level Learning](https://mlsanity.com/paper/2307.10189/) (2023)
- [Capturing Label Distribution: A Case Study in NLI](https://mlsanity.com/paper/2102.06859/) (2021)
- [Distributions In, Distributions Out: The Case for Soft-Label Training](https://mlsanity.com/paper/2511.14117/) (2025)

## Built on by (most cited first, 1 of 1)

- [How Many Humans Are 32 LLM Judges Worth?](https://mlsanity.com/paper/2609.21277/) (2026)

## Abstract (arXiv)

When annotators disagree on a label, the disagreement itself carries signal -- and the number of annotators needed to capture it depends on the evaluation metric. We fine-tune NLI models on label distributions subsampled from ChaosNLI, a dataset providing 100 independent annotator judgments per item, and identify metric-dependent saturation. In our 3-class NLI setting, entropy correlation -- whether the model identifies which items elicit disagreement -- requires N ~ 20-50 annotators to converge, while distributional match (KL divergence) saturates by N ~ 10 (87-95% of improvement across five model seeds). This finding rests on a prior observation: soft labels carry item-specific signal that label smoothing cannot replicate. Across five smoothing intensities, entropy correlation clusters at r ~ 0.45-0.49, while soft labels reach r = 0.643 (p < 0.001); per-item analysis traces this gap to smoothing's inability to distinguish ambiguous items from clear ones. The soft-label advantage replicates across two architectures (DeBERTa, RoBERTa), a non-NLI-pretrained baseline, and an exploratory cross-domain evaluation on content safety. These results suggest that annotation budgets should be informed by the target evaluation metric rather than set uniformly.