Ideas, tested
in the open.
Peer-reviewed publications and preprints across model reliability, multilingual AI, efficient inference, and machine learning fundamentals.
6 publications · 2025—2026
ATMA: Long-Context Language Modeling via Polar Attention and Gated-Delta Compression Memory
We present ATMA, a 378M-parameter hybrid architecture combining Polar Attention with gated-delta recurrent memory for efficient long-context language modeling.
Preprint
Weak-SIGReg: Covariance Regularization for Stable Deep Learning
We introduce a geometric regularizer that enforces covariance isotropy to rescue ViTs and deep MLPs from optimization collapse without architectural hacks.
ICLR 2026 GRaM Workshop
Iterative Layer Pruning for Efficient Translation Inference
This paper describes our submission to The Tenth Conference on Machine Translation (WMT25)
WMT 2025
Language Surgery in Multilingual Large Language Models
How Researchers are Performing Precision Control on Multilingual LLMs. Accepted at Multilingual Representation Learning (MRL) Workshop at ACL 2025
ACL 2025 MRL Workshop
Bemba Speech Translation: Exploring a Low-Resource African Language
This paper describes our submission to the International Conference on Spoken Language Translation (IWSLT 2025)
IWSLT 2025
Early Study: Self-Perturbation Learning for Verifying Mathematical Reasoning
Imagine "2 truth and a lie", but formalized as ML training objective. Accepted at The 10th International Conference on Computer Science and Computational Intelligence (ICCSCI 2025)
ICCSCI 2025