Alex Gulko

Email me at WVd4bGVFQm5kV3

Papers

Undergraduate Honors Research Thesis, The Ohio State University, 2026

FlashStieltjes: Fast and Memory-Efficient Flash-Style Attention with Algebraic Normalization

Alex Gulko · Advised by Sachin Kumar

Softmax is not the only way to do attention. I wrote a flash-style GPU kernel for Stieltjes attention with a 3.37× faster forward pass on an A100 and 4.9× less peak memory than a PyTorch reference of the same attention.

Write-up Knowledge Bank PDF Code

BlackboxNLP Workshop, EMNLP 2025

CE-Bench: Towards a Reliable Contrastive Evaluation Benchmark of Interpretability of Sparse Autoencoders

Alex Gulko*, Yusen Peng*, Sachin Kumar · * equal contribution

Grading how interpretable a sparse autoencoder is with pairs of contrasting stories. No LLM judge needed.

Write-up Paper arXiv Code Dataset

Projects

litl

A lightweight Python framework for writing and benchmarking lossy compressors. You bring one class, litl brings the CLI, the file format, and the metrics.

python data compression research