Source-linked AI summary
Ground-Truth Labels Matter: A Deeper Look into Input-Label Demonstrations
Kang Min Yoo, Junyeob Kim, Hyuhng Joon Kim, Hyunsoo Cho, Hwiyeol Jo, Sang-Woo Lee, Sang-goo Lee, Taeuk Kim
TL;DR
The paper addresses uncertainty about how much input-label correspondence matters in ICL, especially given prior claims of label insensitivity. It introduces quantitative sensitivity metrics and analyzes experimental factors, finding that label mappings have varying effects and that some settings support greater noise resilience.
Problem
Prior findings suggested that input-label correspondence contributes less to ICL performance than expected, despite its central role in supervised learning.
Method
The paper combines Label-Correctness Sensitivity and GLER with statistical analyses of performance under label corruption.
Results
The integrity of input-label mappings is a crucial ICL factor, with its impact varying across experimental configurations and models.
Takeaways & Limitations
Prompt and model choices can help ICL tolerate noisy label demonstrations while maintaining performance.
Takeaways & Limitations
Results using the Huggingface GPT-NeoX implementation may not match the original implementation's performance.
Abstract
from arXiv · showhide
Despite recent explosion of interests in in-context learning, the underlying mechanism and the precise impact of the quality of demonstrations remain elusive. Intuitively, ground-truth labels should have as much impact in in-context learning (ICL) as supervised learning, but recent work reported that the input-label correspondence is significantly less important than previously thought. Intrigued by this counter-intuitive observation, we re-examine the importance of ground-truth labels in in-context learning. With the introduction of two novel metrics, namely Label-Correctness Sensitivity and Ground-truth Label Effect Ratio (GLER), we were able to conduct quantifiable analysis on the impact of ground-truth label demonstrations. Through extensive analyses, we find that the correct input-label mappings can have varying impacts on the downstream in-context learning performances, depending on the experimental configuration. Through additional studies, we identify key components, such as the verbosity of prompt templates and the language model size, as the controlling factor to achieve more noise-resilient ICL.
1 Introduction
The paper revisits claims that input-label correspondence matters little in ICL, finding that ground-truth labels can substantially affect performance under different experimental settings. It introduces quantitative metrics and analyzes factors associated with label-noise sensitivity.
- 1 Introduction: Ground-truth labels changed accuracy by as much as 80% for hate speech on GPT-J, with similar observations on GPT-3 and TREC.These cases motivated re-examination of conclusions based on aggregated results.
- 1 Introduction: The study reports that input-label correspondence can be more important than previously considered, depending on datasets, metrics, and prompt templates.The trend appeared across GPT-J and GPT-3.
- 1 Introduction: The authors propose Label-Correctness Sensitivity and GLER to quantify how ground-truth label demonstrations affect ICL performance.Sensitivity measures performance change under a fixed amount of label corruption.
- 1 Introduction: They analyze inference methods, language models, and advanced ICL strategies to identify conditions associated with sensitivity to noisy demonstrations.The broader motivation is to understand when ICL can remain robust to incorrect labels.
2 Looking Deeper into Ground-Truth Labels
The paper revisits claims that input-label correspondence matters little in ICL, showing that sensitivity varies across tasks and experimental settings. It introduces quantitative metrics and finds that label correctness can substantially affect performance, while task difficulty and ICL method shape observed sensitivity.
- 2.1 Limitations of the Existing Work: The authors identify over-generalization in prior work because mean aggregation obscures individual-task variation and relies on limited experimental settings.A counter-example uses Direct ICL, accuracy, and all 17 datasets instead of the earlier Noisy Channel, macro-F1, and five-dataset configuration.
- 2.1 Limitations of the Existing Work: The paper argues that quantitative analysis is necessary because prior reliance on human judgment cannot measure the degree of input-label sensitivity systematically.The goal is not only to establish whether insensitivity exists, but also to quantify how much performance changes under label corruption.
- 2.3 Metrics for Measuring the Impact of Input-Label Demonstrations: The paper introduces Label-Correctness Sensitivity and GLER to quantify how ground-truth label demonstrations affect downstream ICL performance.Sensitivity uses regression against the percentage of correctly labeled examples; GLER compares ground-truth and random-label performance relative to zero-shot performance.
- 2.4.2 Label Correctness Does Affect Performance: 0.309% accuracy dropped per percentage-point decrease in label correctness in the authors’ setup, with aggregated sensitivity fitting strongly across configurations.The reported regression fits range from 0.81 to 0.86, and sensitivity was lower for the Channel method than for the Direct method.
- 2.4.3 Label Demonstration Impact is Highly Varied Across Tasks and Settings: Sensitivity ranged from small negative values to 0.815 across datasets, showing that label-demonstration impact is highly task-dependent.The authors associate low sensitivity with tasks where models struggle to exploit demonstrations and report that tasks with sensitivity below 0.1 struggle to achieve meaningful classification performance.
3 When Do the Ground-Truth Labels Actually (Not) Matter?
Label sensitivity varies substantially with ICL configuration. Noisy Channel inference and calibration reduce sensitivity while preserving or improving performance, and verbose prompts and larger models further shape noise resilience.
- 3.1 Techniques for In-context Learning: Channel reduces label sensitivity while maintaining similar accuracy across GPT-NeoX and GPT-J.
- 3.1 Techniques for In-context Learning: CBU reduces label sensitivity while generally improving ICL performance on GPT-J and GPT-NeoX.
- 3.2 Prompt Template: More prompt examples increase sensitivity, GLER, and ground-truth-label performance because both final performance and label impact improve.
- 3.2 Prompt Template: Greater template verbosity decreases sensitivity and GLER without deteriorating performance.
- 3.3 Model Sizes: Sensitivity and GLER plateau after GPT-J’s 6B scale, while ground-truth-label performance continues improving with model size.
4 Discussion
Additional analyses show that input-label correspondence remains important even when label-space information is removed, while prompt verbosity and model scale affect sensitivity to corrupted demonstrations.
- 4.1 The Complementary Relationship between Input-label Correspondence and Label-space Demonstrations: Prompt verbosity decreases sensitivity and impact ratio without deteriorating final ground-truth-label performance.
- 4.3 Model Scale: Sensitivity and GLER plateau from GPT-J’s 6B scale, whereas ground-truth-label performance continues improving through larger models.
- 4.1 The Complementary Relationship between Input-label Correspondence and Label-space Demonstrations: Prior-free labels outperform random labels, showing that models capture input-label correspondence without explicit label-space demonstrations.
- 4.1 The Complementary Relationship between Input-label Correspondence and Label-space Demonstrations: Alphabetical and numerical prior-free labels outperform random-token labels, whose word semantics may introduce unintended biases.
- 4.2 Label Distribution: Balanced demonstrations reduce performance and sensitivity relative to data-distribution sampling, yet average sensitivity remains 0.189 for GPT-NeoX and 0.308 for GPT-J.
5 Related Work
ICL offers parameter-free adaptation for large language models, while prior findings on demonstration quality motivate examining how label distributions affect performance.
- ICL adapts large language models to downstream tasks through task-specific prompts without explicit fine-tuning.
- Balanced label demonstrations degrade all metrics on five imbalanced datasets because they create distributional shift.
6 Conclusion and Future Work
The paper concludes that input-label mapping integrity is crucial for ICL and identifies conditions that improve resilience to label corruption.
- GLER and input-label sensitivity quantify how input-label relationships affect ICL performance.
- Ablation studies identify conditions that improve insensitivity to label corruption while maintaining healthy performance.
Limitations
The findings are constrained by prompt sensitivity, incomplete isolation of demonstration factors, model implementation choices, and limited model coverage. The authors mitigate some prompt-related variation by adopting templates from prior work.
- Prompt sensitivity: Prompt choices may affect the findings because prompt engineering across the many datasets and settings was infeasible.The templates and prompts were therefore adopted from well-studied previous works where possible.
- Scope of isolation: The experiments isolate input-label correspondence imperfectly because other demonstration factors, including input-space demonstrations and label distributions, may influence results.The authors note that disentangling diverse demonstration aspects is highly difficult.
- Implementation: The implementation uses Huggingface GPT-NeoX, which may underperform the original implementation.This creates a model-implementation boundary for interpreting results involving GPT-NeoX.
- Model coverage: The main experiments focus on GPT-NeoX 20B and GPT-J 6B, with additional models included in Section 3.3.The listed additional models are GPT2-XL 1.5B, GPT-Neo 2.7B, and GPT-3 175B.
A.2 Full Dataset
The full-dataset appendix describes a broad classification evaluation, its metrics and prompt templates, and the procedure for corrupting demonstration labels. It also specifies baseline and reproducibility choices for the experiments.
- Dataset and metrics: The evaluation covers 17 text-classification datasets spanning sentiment, paraphrase detection, natural-language inference, hate-speech detection, and multiple domains.All datasets are drawn from Huggingface datasets.
- Dataset and metrics: Accuracy is the primary metric, while F1 is discussed as useful for imbalanced classes but less intuitive and dependent on the averaging method.The appendix explains the trade-off between precision and recall underlying F1.
- Templates: The study uses minimal templates by default and additionally evaluates manual and verbose templates in an ablation.Templates are adopted from prior work where possible; CBU uses a manual baseline because minimal templates degraded performance in preliminary tests.
- Label-corruption protocol: Label corruption is evaluated by replacing fixed amounts of correct labels with incorrect ones and comparing downstream task performance.The setup includes no demonstrations, partially correct demonstrations, random labels, shuffled labels, and a majority-class baseline.
B Full Results
The appendix reports full experimental results for the 17-dataset evaluation, including GPT-NeoX results in Tables 4 and 5.
- Full results: Tables 4 and 5 provide full results for the 17-dataset evaluation with GPT-NeoX.The supplied passage identifies the result tables but does not report their individual values.
C More Results on the Sensitivity vs Task Difficulty Plot
The sensitivity-versus-difficulty analysis relates label sensitivity to task difficulty across 17 datasets for Direct and Channel approaches. The figure uses relative performance and distinguishes the two approaches by color.
- Sensitivity versus task difficulty: Figure 11 plots sensitivities for 17 datasets against task difficulty measured by relative performance for accuracy and F-1 scores.The figure compares Direct and Channel approaches using separate trend lines.
- Sensitivity versus task difficulty: The plot colors Direct orange and Channel blue, with a dashed vertical line marking neutral performance relative to random baselines.Best-fit linear lines summarize the general sensitivity trend against task difficulty.
D Label-Correctness Correlation
Label correctness is strongly associated with in-context learning performance across most tasks and metrics, with a macroaverage Pearson correlation of 0.895 across 18 tasks.
- D Label-Correctness Correlation: 0.895 macroaverage Pearson correlation across 18 tasks supports a strong linkage between label correctness and in-context learning performance.Most task-level correlations exceed 0.9 across performance measures, although the macroaverage p-value is 0.057.
- D Label-Correctness Correlation: Strong positive correlations are observed across all tasks and metrics except for outliers.