Are LLMs Better than Reported? Detecting Label Errors and Mitigating Their Effect on Model Performance

TL;DR

Mislabeled data in NLP benchmarks distorts model evaluation and progress. The authors propose an ensemble of zero-shot LLMs to detect label errors by identifying high-confidence disagreements with original labels, followed by expert re-annotation or data cleaning. They demonstrate substantial label-error prevalence (6–21%) across TRUE datasets and SummEval, with higher LLM confidence yielding higher precision in error detection; correcting labels improves training performance by up to 4% and evaluation by up to 15% under accurate labeling. The work highlights scalable, cost-effective data-curation workflows that can tighten benchmark reliability and reshape perceptions of model capability, suggesting LLM-assisted validation as a core component of future NLP evaluation pipelines.

Abstract

NLP benchmarks rely on standardized datasets for training and evaluating models and are crucial for advancing the field. Traditionally, expert annotations ensure high-quality labels; however, the cost of expert annotation does not scale well with the growing demand for larger datasets required by modern models. While crowd-sourcing provides a more scalable solution, it often comes at the expense of annotation precision and consistency. Recent advancements in large language models (LLMs) offer new opportunities to enhance the annotation process, particularly for detecting label errors in existing datasets. In this work, we consider the recent approach of LLM-as-a-judge, leveraging an ensemble of LLMs to flag potentially mislabeled examples. We conduct a case study on four factual consistency datasets from the TRUE benchmark, spanning diverse NLP tasks, and on SummEval, which uses Likert-scale ratings of summary quality across multiple dimensions. We empirically analyze the labeling quality of existing datasets and compare expert, crowd-sourced, and LLM-based annotations in terms of the agreement, label quality, and efficiency, demonstrating the strengths and limitations of each annotation method. Our findings reveal a substantial number of label errors, which, when corrected, induce a significant upward shift in reported model performance. This suggests that many of the LLMs' so-called mistakes are due to label errors rather than genuine model failures. Additionally, we discuss the implications of mislabeled data and propose methods to mitigate them in training to improve performance.

Figures (14)

• Figure 1: An illustration of our approach for detecting and addressing mislabeled data: (1) Re-label examples from existing datasets using an ensemble of LLMs. (2) Identify strong disagreements between the LLM's predictions and the original labels (i.e., high confidence in a different label), flagging examples based on confidence levels. Our findings show that LLMs detect between 6% and 21% of label errors, and higher LLM confidence is strongly associated with improved precision in error detection. (3) In the training set, we either filter or flip flagged examples, leading to an increase of up to 4%. For the test set, flagged examples are re-annotated by experts to make sure the evaluation is accurate. Under accurate evaluation, the performance of LLMs is up to 15% higher.
• Figure 2: When LLMs disagree with original labels - who is correct? (Top) TRUE (Bottom) SummEval. As the LLM's confidence grows, so does the precision of identifying an error in the original labels.
• Figure 3: The power of ensemble. (Top) TRUE (Bottom) SummEval. As the ensemble size increases (x-axis), its performance against gold labels (Left), and its ability to detect label errors (Right) improve.
• Figure 4: Annotation approaches comparison.
• Figure 5: (x-axis) at list $x$ annotations per annotator. (Right y-axis) The number of annotators with at least $x$ annotations (bins). (Left y-axis) the average F1-score or accuracy for all user annotations with at least $x$ annotations.