My research on ML systems efficiency spans three main areas:
developing efficient inference methods, addressing resource constraints
in multilingual settings, and building practical applications for social impact.
ML Efficiency & Systems
Steven Kolawole*, Lucio Dery*, Jean-François Kagy, Virginia Smith, Graham Neubig, Ameet Talwalkar
under review
Presents Bonsai, a forward-pass-only structured pruning method that outperforms gradient-based approaches
while using 3× less memory, making model compression accessible on everyday hardware.
Nnaemeka Obiefuna, Samuel Oyeneye, Similoluwa Odunaiya, Iremide Oyelaja, Steven Kolawole
under review
Comprehensive benchmarking study examining the computational/energy overhead of privacy-preserving techniques
in privacy-sensitive deep learning systems; conducted with independent student researchers at ML Collective.
Steven Kolawole, Keshav Santhanam, Virginia Smith, Pratiksha Thaker
NeurIPS 2025 Datasets & Benchmarks Track
Introduces PARALLELPROMPT, a benchmark revealing that 10% of natural user queries contain latent parallelism.
Demonstrates semantic decomposition methods achieving up to 5× speedups without hardware modifications.
Duncan Soiffer, Steven Kolawole, Virginia Smith
EMNLP 2025 Industry Track
Extends agreement-based cascading to open-ended generation tasks, leveraging meaning-level consensus for cost-effective
routing of language model queries without requiring additional training data or model modifications.
Steven Kolawole*, Don Dennis*, Ameet Talwalkar, Virginia Smith
TMLR 2025
Develops a training-free cascading framework using ensemble agreement as a confidence signal for routing.
Achieves 2-25× cost reductions while maintaining/improving accuracy over diverse tasks.
Resource-Constrained NLP
Mardiyyah Oduwole, Oluwatosin Olajide, Jamiu Suleiman, Faith Hunja, Busayo Awobade, Fatimo Adebanjo, Comfort Akanni, Chinonyelum Igwe, Peace Ododo, Promise Omoigui, Abraham Owodunni, Steven Kolawole
arXiv preprint, 2025
Comprehensive study on data augmentation techniques for African languages, showing that methods like back-translation and
token replacement can significantly improve translation quality in low-resource settings.
Busayo Awobade*, Mardiyyah Oduwole*, Steven Kolawole*
ICLR 2024 (AfricaNLP)
Investigates compression techniques on AfriBERTa, demonstrating that pruning, knowledge distillation,
and quantization remain effective in the "low-resource double-bind" of small-data language models.
Colin Leong, Herumb Shandilya, Bonaventure Dossou, Atnafu Tonja, Joel Mathew, Abdul-Hakeem Omotayo, Oreen Yousuf, Zainab Akinjobi, Chris Emezue, Shamsudeen Muhammad, Steven Kolawole, Younwoo Choi, Tosin Adewumi
ICLR 2023 (AfricaNLP)
Explores efficient training methods for African language processing under computational constraints,
addressing the challenge of limited data and limited compute resources simultaneously.
Applied ML & Social Good
Nahid Alam*, Steven Kolawole*, Simardeep Sethi*, Nishant Bansali, Karina Nguyen
arXiv preprint, 2023
Comprehensive survey examining how Vision Transformers can be optimized for mobile deployment,
analyzing architecture modifications and efficiency techniques for resource-constrained environments.
Steven Kolawole, Opeyemi Osakuade, Nayan Saxena, Babatunde Kazeem Olorisade
IJCAI 2022 AI for Social Good Track
Develops a sign-to-speech system for Nigerian Sign Language to bridge communication gaps, combining computer vision and NLP techniques to translate sign language videos into spoken language.
This work earned a bunch of local awards (including the national AI Champion award from the Nigeria Computer Society), demonstrating practical AI for social impact.
* denotes equal contribution
This list includes peer-reviewed publications, workshop papers, and preprints.
For the most up-to-date list with citation counts, please visit my Google Scholar profile.