Preference benchmarks are built by hiring annotators, and the identity of those annotators is treated as an implementation detail. We measure what that detail buys. On the 2,885 MultiPref items where both pools are internally unanimous, so no tie-breaking convention is consulted at all, expert and crowd annotators assign a different majority label to 23.6% and name the opposite winner on 9.2%; on the 246 comparably unanimous MT-Bench cells, benchmark authors and recruited experts differ on 30.5% and reverse on 8.5%. Yet on both corpora the resulting model leaderboards are bit-identical: Kendall tau = 1.00 with zero of six models displaced. That invariance is far weaker evidence than it looks, and we quantify how weak. Switching pools moves a model's win rate by 1.9pp (SD), one adjacent pair in our own leaderboard sits 0.8pp apart and had a 38% chance of swapping, and an item-level bootstrap displaces at least one model in 28% of resamples. The observed zero is the common outcome, not a property of aggregation: on the same measured perturbation, a ten-model leaderboard is displaced with probability 0.86 and a twenty-model leaderboard with probability 0.9997. Reporting a six-model leaderboard is safe; the safety does not generalise, and everything that consumes labels per item is not safe at any size. We make the distinction precise, show that a widely used dataset's stated assumption of no intra-group annotator variability is false, and show that an LLM judge tracks the crowd pool over the expert pool on all three models we test, including one from a different vendor. All code, per-call outputs, and pre-registered decision rules will be released upon acceptance.
Hadi Mohammadi, Tina Shahedi, Robert A. Bagheri +2cs.CL cs.CY cs.LG
When people label text for sexism, they often disagree, and not because some of them are wrong: they genuinely perceive sexism differently. Most NLP systems discard this disagreement by collapsing it into a majority vote. We propose the Multi-Agent Perspectivist Preference Optimization (MAP-PO) framework to keep these different perspectives. On the EXIST 2024 dataset of labeled English and Spanish tweets, we first cluster annotators by their labeling behavior rather than their demographic attributes. We then fine-tune one Large Language Model agent per cluster to reproduce that cluster's annotation behavior, and coordinate the agents with preference optimization that combines individual and team-level rewards. We evaluate MAP-PO in four settings defined by two languages and two backbone language models, asking whether each agent reproduces the annotations of its own cluster and whether the agents together reproduce the majority label. Two findings hold in all four settings. First, without fine-tuning the agents behave almost identically, so cluster-specific training is necessary. Second, we show that training each agent only on the labels of its own cluster pushes the agents far beyond the clusters they should represent, while adding a shared team-level training signal consistently keeps each agent calibrated to its cluster.
Subjective NLP tasks often exhibit systematic annotator disagreement, requiring models that represent uncertainty rather than collapse it. We introduce Ensemble Diversity Optimization (EDO), a prediction-space framework that jointly optimizes ensemble weights, effective cardinality, and calibration through a unified differentiable objective. EDO learns ensemble composition and size end-to-end via Gumbel-Softmax relaxation and incorporates a signed diversity regularizer, tuned on validation data, to steer optimization toward either preserving or suppressing disagreement. This regularization prevents ensemble collapse and enables controlled navigation of the utility-calibration trade-off. The framework integrates a soft F1 surrogate, class-weighted cross-entropy to address imbalance, and reliability-weighted diversity to regulate intra-ensemble variability. Experiments on four subjective text-classification benchmarks (ArMIS, ConvAbuse, HS-Brexit, MD-Agreement) show that EDO substantially improves probabilistic calibration, reducing cross-entropy (40-78% depending on baseline) and lowering Brier scores relative to Soft-CE, Soft-MD, Top-5 Voting, and WEL, while maintaining competitive F1 and better alignment with annotator distributions. These results demonstrate that jointly optimizing ensemble structure with a signed diversity regularizer provides an efficient, model-agnostic approach for modeling human subjectivity in supervised learning.
Hate speech annotation pipelines routinely collapse annotator disagreement into majority vote labels before training. We show that this aggregation is not neutral: 42.6% of all annotator disagreement in HateXplain concentrates specifically at the hate/offensive boundary, a pattern consistent with annotators applying different thresholds for where hate begins (chi-squared = 135.199, df = 2, p < 0.0001). Both a hard-label BERT model (Model A) and a soft-label model (Model B) drop 22 percentage points in accuracy from agreed posts (~80%) to disagreement posts (~58%), confirmed at p < 0.0001. A per-annotator multi-head model (Model C) widens this gap further to 28 points while collapsing offensive disagreement accuracy to 0.245. Critically, Model A expresses significantly higher confidence on boundary case errors than Model C (0.710 vs. 0.495, p < 0.0001), meaning standard evaluation metrics will not detect the failure. Three downstream interventions of increasing sophistication all fail to recover boundary accuracy. We argue the problem is structural. Majority vote presents a contested judgment as ground truth, and models inherit that false certainty. The intervention must be upstream in annotation design.
Aleksander Szczęsny, Wiktoria Mieleszczenko-Kowszewicz, Maciej Markiewicz +5cs.CL cs.AI
IMPACTeen is a dataset of textual social influence scenarios spanning interpersonal, media-based, and digital settings in an adolescent context. It contains 1,021 texts, 5,100 individual annotation records, and gold labels for social influence techniques, with each text annotated from five distinct perspectives: teenagers, parents, psychologists, communication experts, and teachers. The resource was constructed through constrained LLM generation, followed by a two-step human editing and validation phase aimed at ensuring youth-context realism. A multi-dimensional annotation covered influence presence, techniques, intentions, consequences, resistance, reactions, and annotation confidence. The dataset supports research on social influence detection, annotator disagreement, cross-lingual modeling, and the training and evaluation of language models. The dataset was created in Polish and is accompanied by a corresponding English version.
Many human-centered tasks, including natural language inference (NLI) and emotion recognition (ER), have multiple plausible interpretations, leading to label ambiguity and challenging disagreements across human annotators. As LLMs are increasingly deployed in real-world settings, faithfully modeling such ambiguity is essential to identify contested inputs, preserve variability in ambiguous cases, and capture the full distribution of human judgments. Yet, existing LLM alignment approaches have predominantly assumed a single correct label, excluding annotator disagreement during optimization. Instead of treating this ambiguity as noise, we show how to treat it as information that improves model behavior through a new algorithm called SMARTLY HANDLING AMBIGUOUS LABELS IN ALIGNING LLMS (SHALA-LLM). This reinforcement learning framework provides a new way for LLMs to learn directly from annotator distributions while dynamically prioritizing highly ambiguous samples during optimization. Experiments on ambiguity-sensitive NLI and ER benchmarks, including ChaosNLI, GoEmotions, and MSP-Podcast, demonstrate that SHALA-LLM improves agreement with annotator label distributions, e.g. on ChaosNLI, it reduces Jensen-Shannon Distance by up to 62.1%. At the same time, SHALA-LLM improves F1 by up to 16.7%, showing that modeling annotator disagreement can also strengthen classification performance.
When annotators disagree, that disagreement can reflect epistemic uncertainty rather than simple label noise. We study hard-label delivery as an alternative to the usual choices of collapsing votes to a single label or training directly on the empirical soft-label distribution. We focus on two primary hard-label methods: multipass, which cycles through observed votes while keeping the dataset size fixed, and stochastic label sampling (SLS), which samples one label per example at the start of each epoch. On CIFAR-10H, we find that when only a small number of annotations per example is available, hard-label delivery improves over soft-label training, with larger improvements where the sparse empirical target is farther from the full annotator distribution. When full annotator distributions are available, both hard-label methods match soft-label training. We use deterministic control as an ablation of multipass and shuffled SLS as a control that breaks the example-to-distribution match. We also show that SLS and soft-label cross-entropy optimize the same expected objective. Hard-label delivery also converges to flatter basins, with supporting descriptive evidence from OOD detection on SVHN and CIFAR-100. Overall, these results suggest that multipass is a strong practical default when raw vote counts are available, while SLS offers a lightweight alternative that remains competitive when only a few votes per example are available and matches soft-label training when full annotator distributions are available.
Human evaluation remains the primary standard for assessing modern AI systems, yet annotator disagreement, bias, and variability make system rankings fragile under standard majority vote aggregation. Majority vote discards annotator reliability and item-level ambiguity, often yielding unstable comparisons across annotator subsets. We introduce STABLEVAL, a disagreement-aware evaluation framework that models latent item correctness and annotator-specific confusion patterns to produce posterior expected item credit and calibrated agent-level scores. Unlike label-denoising approaches such as Dawid-Skene, STABLEVAL is explicitly designed for stable and uncertainty-aware system evaluation rather than hard label recovery. We formalize ranking stability as a first-class evaluation objective and analyze how aggregation methods preserve or distort underlying annotator behavior. Across controlled synthetic experiments and multiple real-world human-annotated benchmarks, majority vote exhibits increasing score error and ranking instability under annotator heterogeneity and adversarial noise, while STABLEVAL yields more stable and statistically grounded system rankings. These results demonstrate that modeling disagreement is essential for robust and reproducible AI evaluation.