Steven Kolawole

Doctoral Researcher, Carnegie Mellon University
Language Technologies Institute, School of Computer Science
Email: skolawol[at]cs[dot]cmu[dot]edu
Steven Kolawole

Hi 馃憢馃徔 I'm a PhD student at CMU, grateful to be advised by Virginia Smith. My research is on efficient inference, which for me means making ML models cheaper/faster to run, from serving clusters down to devices with a few megabytes of memory. So far, that has spanned model compression (Bonsai, EdgeTRM), inference-time memory (EpiKV), adaptive inference (ABC, Semantic Agreement, speculative decoding), and semantic parallelism within requests (ParallelPrompt).

Lately I've been thinking about where inference should happen, and the constraints that come with each choice. Once models start acting for people as agents instead of just answering them, a lot of that work has to run on the device, near the data, for both latency and privacy reasons. I want to know what it takes for a model to reason well there: on the hardware it has, remembering what it can afford to, and knowing when it's out of its depth.

I spent summer 2026 on the inference optimization science team at AWS AGI Foundations, working with Nathan Pemberton and Kyle Ulrich on speculative decoding.

Selected Publications

Model 路 compression
Everybody Prune Now: Structured Pruning of LLMs with only Forward Passes (Bonsai)
Steven Kolawole*, Lucio Dery*, Jean-Fran莽ois Kagy, Virginia Smith, Graham Neubig, Ameet Talwalkar
arXiv, 2024
what it does
Prunes LLMs using only forward passes, instead of the gradients backprop needs. It still matches gradient-based methods, and prunes 7-8B models on a single commodity GPU.
arXiv
Memory 路 KV cache
Epiphany-Aware KV Cache Eviction Without the Attention Matrix (EpiKV)
Steven Kolawole, Virginia Smith
preprint, 2026
what it does
Decides what to evict from the KV cache by how much a token's representation changes, not by its attention weights. So it never builds the attention matrix, runs inside FlashAttention stacks, and fits several times more context.
arXiv
Serving 路 adaptive routing
Agreement-Based Cascading for Efficient Inference (ABC)
Steven Kolawole*, Don Dennis*, Ameet Talwalkar, Virginia Smith
TMLR, 2025
what it does
Stacks models from cheap to expensive and only defers to a bigger one when the cheap ones disagree. No confidence scores, no trained router, and no retraining for each new task or data distribution. It drops into an existing setup and cuts serving cost several-fold at the same accuracy.
OpenReview arXiv
Workload 路 parallelism
ParallelPrompt: Extracting Parallelism from Large Language Model Queries
Steven Kolawole, Keshav Santhanam, Virginia Smith, Pratiksha Thaker
NeurIPS 2025, Datasets & Benchmarks Track
what it does
Lots of real prompts contain sub-tasks that don't depend on each other, but systems still run them one by one. I built the tooling (a C++ engine, schema extraction, vLLM/SGLang hooks) to run them in parallel: up to 7脳 faster while keeping >90% of output quality.
OpenReview arXiv
Complete publication list

News

More updates

Community

I lead ML Collective Africa, a pan-African sub-community of ML Collective, where we back under-served, early-career researchers with research pods, mentorship, training, and compute, so they can do good research wherever they're starting from. Some have gone on to top venues and good grad programs. The goal is extending accessibility such that people who don't have a lab still don't face too many constraints in contributing to global research. My efficiency research is one part of that; this community is the other part. Since 2022 I've run annual fundraisers that send African students to the Deep Learning Indaba (roughly $50k raised so far), and I occasionally contribute to Black in AI's ELAI program.

From 2023 to 2025, I mentored 20+ underrepresented graduate-school aspirants a year at STEM for Development. During undergrad, across the grassroots tech communities I led, I helped run dozens of technical training programs that reached "a bunch" of students, and personally taught several hundred of them.

Outside research and community work, I enjoy powerlifting, amateur boxing, watching Liverpool FC matches, and reading widely. Email me if you want to talk about any of this.

The Four Six Years of BSc. if you're curious :)
Steven speaking at PyCon Italia
Speaking at PyCon Italia, one of the one million and one conferences. :)

Before CMU, I did my BSc in Computer Science at the Federal University of Agriculture, Abeokuta (FUNAAB), Nigeria, advised by Adebayo Abayomi-Alli. Academic-union strikes and the COVID lockdown stretched a four-year degree into six. Fortunately, that gave me far more room than a typical undergrad to figure out what I actually cared about.

I used it. Around 27 tech conferences across the globe, a pile of hackathons and internships, and, from early 2021, learning to do research independently with ML Collective (with cameos at Masakhane and Cohere For AI), mentored by Rosanne Liu and Jason Yosinski, who took me on when I had little to show. That was the luckiest thing that happened to me in those years. My first finished project (on sign-language understanding) involved building the first large-scale NSL dataset; the work got into a NeurIPS workshop, then got published at IJCAI, and then won the Nigeria Computer Society's National AI Champion award. Later, friends and I built a real-time opinion-mining system for digital assets and won a ~$115K Algorand Foundation grant to ship it, which let me keep doing independent research without worrying much about rent. :)

I'm a "community-taught" ML practitioner, so a lot of those years went to giving back: Data Science Nigeria, Google Developer Student Clubs, She Code Africa, NACOS, and ML Collective. That's also where my research taste comes from: I learned ML scrounging on compute that couldn't run most of my experiments, and ML Collective Africa is the research group people like me didn't have.

For the human record: in the earliest years I directed choirs at local churches (vocals, drums, piano); my BSc was also roundly punctuated by real existential crises [1, 2]; and my final year saw me trade a long, award-worthy run of social awkwardness for a reputation as the still-weird-but-now-funny guy, obsessively fine-tuning my Afro-pop dance moves and stepping into a spotlight I'd spent years too awkward for.