Dissecting Subjectivity and the "Ground Truth" Illusion in Data Annotation

In machine learning, "ground truth" refers to the assumed correct labels used to train and evaluate models. However, the foundational "ground truth" paradigm rests on a positivistic fallacy that treats human disagreement as technical noise rather than a vital sociotechnical signal. This systematic literature review analyzes research published between 2020 and 2025 across seven premier venues: ACL, AIES, CHI, CSCW, EAAMO, FAccT, and NeurIPS, investigating the mechanisms in data annotation practices that facilitate this "consensus trap". Our identification phase captured 30,897 records, which were refined via a tiered keyword filtration schema to a high-recall corpus of 3,042 records for manual screening, resulting in a final included corpus of 346 papers for qualitative synthesis. Our reflexive thematic analysis reveals that systemic failures in positional legibility, combined with the recent architectural shift toward human-as-verifier models, specifically the reliance on model-mediated annotations, introduce deep-seated anchoring bias and effectively remove human voices from the loop. We further demonstrate how geographic hegemony imposes Western norms as universal benchmarks, often enforced by the performative alignment of precarious data workers who prioritize requester compliance over honest subjectivity to avoid economic penalties. Critiquing the "noisy sensor" fallacy, where statistical models misdiagnose cultural pluralism as random error, we argue for reclaiming disagreement as a high-fidelity signal essential for building culturally competent models. To address these systemic tensions, we propose a roadmap for pluralistic annotation infrastructures that shift the objective from discovering a singular "right" answer to mapping the diversity of human experience.

VeriFact: Enhancing Long-Form Factuality Evaluation with Refined Fact Extraction and Reference Facts

Large language models (LLMs) excel at generating long-form responses, but evaluating their factuality remains challenging due to complex inter-sentence dependencies within the generated facts. Prior solutions predominantly follow a decompose-decontextualize-verify pipeline but often fail to capture essential context and miss key relational facts. In this paper, we introduce VeriFact, a factuality evaluation framework designed to enhance fact extraction by identifying and resolving incomplete and missing facts to support more accurate verification results. Moreover, we introduce FactRBench , a benchmark that evaluates both precision and recall in long-form model responses, whereas prior work primarily focuses on precision. FactRBench provides reference fact sets from advanced LLMs and human-written answers, enabling recall assessment. Empirical evaluations show that VeriFact significantly enhances fact completeness and preserves complex facts with critical relational information, resulting in more accurate factuality evaluation. Benchmarking various open- and close-weight LLMs on FactRBench indicate that larger models within same model family improve precision and recall, but high precision does not always correlate with high recall, underscoring the importance of comprehensive factuality assessment.

FactBench: A Dynamic Benchmark for In-the-Wild Language Model Factuality Evaluation

Language models (LMs) are widely used by an increasing number of users, underscoring the challenge of maintaining factuality across a broad range of topics. We first present VERIFY (Verification and Evidence RetrIeval for FactualitY evaluation), a pipeline to evaluate LMs' factuality in real-world user interactions. VERIFY considers the verifiability of LM-generated content and categorizes content units as supported, unsupported, or undecidable based on the retrieved evidence from the Web. Importantly, factuality judgment by VERIFY correlates better with human evaluations than existing methods. Using VERIFY, we identify "hallucination prompts" across diverse topics, i.e., those eliciting the highest rates of incorrect and inconclusive LM responses. These prompts form FactBench, a dataset of 1K prompts across 150 fine-grained topics. Our dataset captures emerging factuality challenges in real-world LM interactions and can be regularly updated with new prompts. We benchmark widely-used LMs from GPT, Gemini, and Llama3.1 family on FactBench, yielding the following key findings: (i) Proprietary models exhibit better factuality, with performance declining from Easy to Hard hallucination prompts. (ii) Llama3.1-405B-Instruct shows comparable or lower factual accuracy than Llama3.1-70B-Instruct across all evaluation methods due to its higher subjectivity that leads to more content labeled as undecidable. (iii) Gemini1.5-Pro shows a significantly higher refusal rate, with over-refusal in 25% of cases. Our code and data are publicly available at https://huggingface.co/spaces/launch/factbench.