Maja Pavlovic

Maja Pavlovic, Silviu Paun, Massimo Poesio

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.

Yuzheng Xu, Annya Dahmani, Matthew D. Blanchard et al.

Human-AI complementarity, the idea that combining human and AI judgments can outperform either alone, offers a promising pathway toward robust oversight of advanced AI systems. However, whether human-AI complementarity can be achieved on realistic tasks remains an open question. We investigate this through two approaches: hybridization and two AI assistance methods (top-2 assistance and subtask delegation), evaluated on a multi-domain dataset of 1,886 samples spanning knowledge, factuality, long-context reasoning, and deception detection. We find only modest complementarity gains. Baseline hybridization yields just +0.4 percentage points (pp) over AI alone (69.3\% vs 68.9\%), limited both by a small complementarity region (only 8.9\% of items where AI errs but humans do not) and the inability of confidence-based routing to identify it, since the model's confidence is similarly distributed across correct and incorrect predictions. Applied when AI has low confidence, top-2 assistance increases human accuracy from 28.4\% to 38.3\%, surpassing AI alone (37.7\%) -- but primarily because humans adopt correct AI suggestions, not because they successfully override AI errors. These findings suggest that the primary bottleneck is not human task accuracy per se, but the ability to route decisions to humans when it matters and to design assistance methods that enable humans to catch AI mistakes. Our quantitative and qualitative analyses pinpoint where and why each method succeeds or fails, offering concrete targets for future work. We will release our dataset and code upon request to support progress toward more effective human-AI collaboration for AI oversight.

Maja Pavlovic, Massimo Poesio

Large Language Models (LLMs) have emerged as powerful support tools across various natural language tasks and a range of application domains. Recent studies focus on exploring their capabilities for data annotation. This paper provides a comparative overview of twelve studies investigating the potential of LLMs in labelling data. While the models demonstrate promising cost and time-saving benefits, there exist considerable limitations, such as representativeness, bias, sensitivity to prompt variations and English language preference. Leveraging insights from these studies, our empirical analysis further examines the alignment between human and GPT-generated opinion distributions across four subjective datasets. In contrast to the studies examining representation, our methodology directly obtains the opinion distribution from GPT. Our analysis thereby supports the minority of studies that are considering diverse perspectives when evaluating data annotation tasks and highlights the need for further research in this direction.

Maja Pavlovic

To be considered reliable, a model must be calibrated so that its confidence in each decision closely reflects its true outcome. In this blogpost we'll take a look at the most commonly used definition for calibration and then dive into a frequently used evaluation measure for model calibration. We'll then cover some of the drawbacks of this measure and how these surfaced the need for additional notions of calibration, which require their own new evaluation measures. This post is not intended to be an in-depth dissection of all works on calibration, nor does it focus on how to calibrate models. Instead, it is meant to provide a gentle introduction to the different notions and their evaluation measures as well as to re-highlight some issues with a measure that is still widely used to evaluate calibration.

Elisa Leonardelli, Silvia Casola, Siyao Peng et al.

Many researchers have reached the conclusion that AI models should be trained to be aware of the possibility of variation and disagreement in human judgments, and evaluated as per their ability to recognize such variation. The LEWIDI series of shared tasks on Learning With Disagreements was established to promote this approach to training and evaluating AI models, by making suitable datasets more accessible and by developing evaluation methods. The third edition of the task builds on this goal by extending the LEWIDI benchmark to four datasets spanning paraphrase identification, irony detection, sarcasm detection, and natural language inference, with labeling schemes that include not only categorical judgments as in previous editions, but ordinal judgments as well. Another novelty is that we adopt two complementary paradigms to evaluate disagreement-aware systems: the soft-label approach, in which models predict population-level distributions of judgments, and the perspectivist approach, in which models predict the interpretations of individual annotators. Crucially, we moved beyond standard metrics such as cross-entropy, and tested new evaluation metrics for the two paradigms. The task attracted diverse participation, and the results provide insights into the strengths and limitations of methods to modeling variation. Together, these contributions strengthen LEWIDI as a framework and provide new resources, benchmarks, and findings to support the development of disagreement-aware technologies.

Maja Pavlovic, Massimo Poesio

With the increasing capabilities of LLMs, recent studies focus on understanding whose opinions are represented by them and how to effectively extract aligned opinion distributions. We conducted an empirical analysis of three straightforward methods for obtaining distributions and evaluated the results across a variety of metrics. Our findings suggest that sampling and log-probability approaches with simple parameter adjustments can return better aligned outputs in subjective tasks compared to direct prompting. Yet, assuming models reflect human opinions may be limiting, highlighting the need for further research on how human subjectivity affects model uncertainty.