Zeyneb N. Kaya
Hi! I am Zeyneb, a student at Stanford University. I work on understanding and pushing the limits of AI, exploring robustness, learning from data (efficiently), and statistics/optimization---among other things.
Most recently, I've worked on physics-based foundation models for computational design as co-founder/CTO @ Topological (YC S25); decentralized AI, synthetic data, and midtraining @ Dria; and RL/reasoning with text diffusion LLMs @ Stanford.
I’m always eager to discuss interesting ideas and opportunities—please reach out!
zeynebnk [at] stanford [dot] edu
Research.
My work aims to advance our understanding of AI and its capabilities, and use that to improve them and push their limits in their fundamental challenges. I'm interested in robustness, data/efficiency, and generalizability in distribution shifts, working in machine learning, statistics, and physics.
Listed below are selected relevant publications.
Semantic Anchoring in Large Language Models: Thresholds, Transfer, and Geometry
Edward Y. Chang, Zeyneb N. Kaya, Ethan Chang
Under Review
Vector Space Distance as a Measurement of Word Embedding Variability in Low-Resource Linguistic Environments
Annie K. Lamar, Zeyneb N. Kaya, Nichole M. Nomura
Under Review
Measuring the Impact of Data Augmentation Methods for Extremely Low-Resource NMT
Zeyneb N. Kaya, Annie K. Lamar
Proceedings of the Sixth Workshop on Technologies for Machine Translation of Low-Resource Languages (LoResMT) @ EACL, 2023
MADLIBS: A Novel Multilingual Data Augmentation Algorithm for Low-Resource Neural Machine Translation
Zeyneb N. Kaya
Regeneron Science Talent Search, 2024 & National Junior Science and Humanities Symposium, 2023
Zeyneb N. Kaya, Souvick Ghosh
arXiv preprint
Full Scope Word Embedding Variability for Low-Resource Languages
Zeyneb N. Kaya, Annie K. Lamar
IEEE MIT URTC, 2023
Zeyneb N. Kaya
Proceedings of the Linguistic Society of America (PLSA), 2023
Women in the Workplace: Analyzing Gender Biases in Corporate Email Communications
Zeyneb N. Kaya
International Conference on Computational Social Science (IC2S2), 2023
What You Say Is What You Think: An Analysis Of Intellectual Humility In Online Discussion Forums
Zeyneb N. Kaya, Manya Sriram
University of California, Santa Barbara, 2022
Ahmet C. Genc, Zeyneb N. Kaya, et al
Annals of the Rheumatic Diseases, 2021
Awards & Recognition.
Etched x Mercor x Cognition Hackathon – 1st Place/$40K Winner 2025
Regeneron Science Talent Search Winner – 5th Place/$90K Winner 2024
Coca Cola Scholar – 2024
PearVC x Anthropic Hackathon – 1st Place/Most Technical Winner, 2025
TreeHacks Scrapybara Prize – 1st Place/$16K-valued Winner, 2025
Geoguessr – Master Tier Player, 2025
Olympiad in Linguistics (Onling) – 10th Place / 1st in USA, 2023
North American Computational Linguistics Olympiad (NACLO) – Finalist, 2023
International Olympiad in Artificial Intelligence (IOAI) – Team USA invited representative (did not attend due to conflicts)
Education.
Stanford University
Computer Science (AI) /
Saratoga High School
+ Dual Enrolled West Valley College
AI Club Co-President. Linguistics Club Founder + President, Chinese Club Events Coordinator.
Dual Enrollment: Differential Equations, Linear Algebra, Multivariable Calculus, Cultural Anthropology
Projects.
MADLIBS
LLaDA-R1
SHIELD.
In-Context Learning of Transformers: A Statistical Mechanics Lens
Linguistic Reasoning: Dissociating Language and Logic
Language Models (can be)
Few-Shot Fakers
NeuroPilot
Designed Multilingual Augmentation of Data with Alignment-Based Substitution, an efficient multilingual synthetic data generation algorithm achieving SOTA performance with less data.
@ Regeneron Science Talent Search 2024
Created LLaDA-R1, a diffusion LLM optimized for reasoning and efficiency at inference time with SFT+RL for dynamic diffusion step adaptation and remasking refinement.
@ Mercor x Etched x Cognition Inference-Time Compute Hackathon 2025
Built SHIELD., a multi-agent RL + tool use framework for automatic identification and remediation of system vulnerabilities. @ Pear VC x Anthropic Hackathon 2025
Investigated statistical physics models explaining in-context learning; applying spin glasses, random matrix theory, and phase transitions towards transformer interpretability.
@ APPPHYS 229 2025
Developed parallel symbol tuning, an approach to improve in-context linguistic reasoning capabilities of LLMs for few-shot language learning.
@ CS 224N 2025
Investigated CoT faithfulness & the role of memorization; Implemented corrupted CoT RL approach.
@ Anthropic Alignment Research Hackathon 2025
Built brain-computer-interface and agentic AI system for brain-powered natural language commands for hands-free computer control.
@ TreeHacks 2025