Publications/Efficient Machine Learning: Data Selection, Pruning, Active & Federated Learning
Conference · 2026
Language Model-Driven Data Pruning Enables Efficient Active Learning
Abdul Hameed Azeemi, Ihsan Ayyub Qazi, Agha Ali Raza
Findings of EACL 2026, Rabat, Morocco · CORE A
Summary
Active learning saves labeling effort by choosing only the most informative examples to annotate, but scoring a huge pool of unlabeled data to find them is itself expensive. ActivePrune is a plug-and-play method that first shrinks the pool: a fast pass using a simple n-gram language model, then a higher-quality selection using a compact LLM, with a reweighting step that brings forward underrepresented examples. Across translation, sentiment analysis, topic classification, and summarization, it outperforms existing data pruning methods and cuts end-to-end active learning time by up to 74% compared with other LLM-based pruning methods.
Cite this paper
Abdul Hameed Azeemi, Ihsan Ayyub Qazi, Agha Ali Raza. Language Model-Driven Data Pruning Enables Efficient Active Learning. Findings of EACL 2026, Rabat, Morocco.
@inproceedings{azeemi2026language,
author = {Abdul Hameed Azeemi and Ihsan Ayyub Qazi and Agha Ali Raza},
title = {{Language Model-Driven Data Pruning Enables Efficient Active Learning}},
booktitle = {Findings of EACL 2026, Rabat, Morocco},
year = {2026},
url = {https://aclanthology.org/2026.findings-eacl.229/}
}
- To Label or Not to Label: Hybrid Active Learning for Neural Machine TranslationCOLING 2025, Abu Dhabi, UAE
- A Survey on Data Selection for Efficient Speech ProcessingIEEE Access, vol. 13, 2025
- Data Pruning for Efficient Model Pruning in Neural Machine TranslationFindings of EMNLP 2023, Singapore
- Representative Subset Selection for Efficient Fine-Tuning in Self-Supervised Speech RecognitionNeurIPS 2023 Workshop (ENLSP), New Orleans, USA
- Learning Fast and Slow: Towards Inclusive Federated LearningECML PKDD 2023, Turin, Italy