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..
KAIROS: Scalable Model-Agnostic Data Valuation
Authors: Jiongli Zhu, Parjanya Prashant, Alex Cloninger, Babak Salimi
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical evaluations on noise, mislabeling, and poisoning benchmarks show that KAIROS consistently outperforms state-of-the-art baselines in both accuracy and runtime. On Image Net, KAIROS achieves up to 15 speedup over the fastest baseline while maintaining superior data valuation quality. |
| Researcher Affiliation | Academia | Jiongli Zhu*, Parjanya Prajakta Prashant*, Alex Cloninger, Babak Salimi University of California San Diego EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 provides the detailed algorithm for computing the influence in the online setting. Algorithm 1 Online update for influence |
| Open Source Code | Yes | Code is available at Github. |
| Open Datasets | Yes | We evaluate on four widely used datasets, including CIFAR-10 [34], STL-10 [10], IMDB [43], and AG News [77], to cover both image and text modalities. |
| Dataset Splits | Yes | In most experiments, we simulate limited clean-data availability by using 10000 noisy training examples and 300 clean validation examples, with a held-out test set of 10000 clean samples. For the smaller STL-10 dataset, we scale down to 3700 training, 300 validation, and 1000 test examples. |
| Hardware Specification | Yes | We extract features with a Res Net-50 encoder and run all methods on a single NVIDIA A100 (40GB VRAM). |
| Software Dependencies | No | The paper mentions using "Open Data Val benchmark" and specific models like ResNet-50 encoder, but does not specify software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x, CUDA 11.x). |
| Experiment Setup | Yes | For KAIROS, we set the Gaussian kernel bandwidth to the median of all pairwise distances and fix the balancing factor in Equation (12) to 0.03. See details in Appendix D. Experiments are repeated five times with different random seeds, and we report the mean (colored regions denote the standard deviations). |