Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Cost-aware LLM-based Online Dataset Annotation
Authors: Eray Can Elumar, Cem Tekin, Osman Yagan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical evaluation on the MMLU and IMDB Movie Review datasets demonstrates that Ca MVo achieves comparable or superior accuracy to full majority voting while significantly reducing labeling costs. |
| Researcher Affiliation | Academia | Eray Can Elumar Dept. of Electrical and Computer Eng. Carnegie Mellon University Pittsburgh, PA 15213 EMAIL Cem Tekin Dept. of Electrical and Electronics Eng. Bilkent University Ankara, Türkiye EMAIL Osman Ya gan Dept. of Electrical and Computer Eng. Carnegie Mellon University Pittsburgh, PA 15213 EMAIL |
| Pseudocode | Yes | The pseudo-code of Ca MVo is provided in Algorithm 1, and consists of six main steps described below. Algorithm 1 Cost-aware Majority Voting (Ca MVo) Algorithm |
| Open Source Code | No | While we do not directly provide data or code, our experimental results can easily be reproduced. We have provided the pseudo-code of our algorithm in Algorithm 1, and we have provided details on the experimental setting in our Experiments Section ( 4), Appendices D and E. The datasets we used and the LLMs we queried are publicly available. Based on these, we believe our results can easily be reproduced. |
| Open Datasets | Yes | Our empirical evaluation on the MMLU and IMDB Movie Review datasets demonstrates that Ca MVo achieves comparable or superior accuracy to full majority voting while significantly reducing labeling costs. We first evaluate Ca MVo on the MMLU dataset Hendrycks et al. [2021a,b] We next test Ca MVo on the IMDB Movie Reviews dataset Maas et al. [2011] We further evaluate Ca MVo on the AG News Classification Dataset Zhang et al. [2015] |
| Dataset Splits | Yes | To reduce computational cost, we restrict our evaluation to the test split, which contains 14,042 instances. We next test Ca MVo on the IMDB Movie Reviews dataset Maas et al. [2011], a balanced binary-sentiment benchmark of 50,000 movie reviews. From the original training set of 120, 000 samples, we uniformly sample 50, 000 instances for our experiments. |
| Hardware Specification | No | LLMs were queried using online LLM providers. The experiments (apart from querying LLMs) do not consume much computer resources, and were performed with a personal computer. |
| Software Dependencies | No | To extract contextual embeddings for Ca MVo, we use the 384-dimensional sentence transformer all-Mini LM-L6-v2 Wang et al. [2020]. |
| Experiment Setup | Yes | The algorithm is configured with α = 0.25, λR = 1, and λL = 1. All models are queried using temperature 0.35 and top-p = 1, where applicable. Ca MVo s hyperparameters are α = 0.7, λR = 5, and λL = 1 |