Devan ShahI am a senior in Computer Science at Princeton University where I am grateful to be advised by Prof. Elad Hazan. Through the Hazan Lab, I work on language modeling and dynamical systems. Separately, I have also worked on projects with Jane Street and TikTok. Broadly, I am interested in better understanding language models, such as improving how models reason, designing post-transformer architectures, and improving optimizers. At Princeton, I help host NLP & CV reading groups, chair our ACM chapter, and am a member of Phi Beta Kappa, Sigma Xi, and Tau Beta Pi. Some recent work I am especially proud of is [1] and [3]. Email / GitHub / Google Scholar / LinkedIn |
|
Work Experience |
|
Princeton Hazan LabSeptember 2024 — PresentI work on language modeling and dynamical systems in the Hazan Lab at Princeton, led by Prof. Elad Hazan. I've worked on projects aiming to accelerate the Spectral Transformer and better model linear dynamical systems, such as such as SpectraLDS, FutureFill, and Google Deluca. |
|
Quantitative Research at Jane StreetMay 2025 — August 2025I interned as a Quantitative Research intern at Jane Street. |
|
Recommendation Systems at TikTokMay 2024 — August 2024As a machine learning engineering intern, I improved the recommendation system for ecommerce videos through projects focusing on improving multi-interest modeling. |
Show 3 more experiences
|
EEG Vision Embeddings at CareYayaNovember 2023 — May 2024I studied the reconstruction of viewed images from brain activity, improving performance by conditioning on a user's camera roll. We aimed to better align EEG embeddings with CLIP space. |
|
Scooter Authentication at Cal Poly PomonaJune 2023 - April 2024I worked on continuously authentication of mobility scooter riders based on their posture patterns. We worked on developing embeddings based on motion that could be suitable for identification or auth. |
|
Price Prediction at Ticket Wallet (YC X25)June 2023 — August 2023I worked at an early-stage startup to advise and design data pipelines, pricing algorithms, and refine product pitches. I also worked on algorithms to determine if a ticket was likely to be sold. |
Research |
|
SpectraLDS: Provable Distillation for Linear Dynamical SystemsDevan Shah, Shlomo Fortgang, Sofiia Druchyna, Elad Hazan NeurIPS 2025 website / paper / code / poster / slides |
|
UpSkill: Mutual Information Skill Learning for Structured Response Diversity in LLMsDevan Shah*, Owen Yang*, Daniel Yang, Chongyi Zheng, Benjamin Eysenbach NeurIPS 2025 Workshop on Scaling Environment for Agents website / paper / code / poster |
|
FutureFill: Fast Generation from Convolutional Sequence ModelsNaman Agarwal, Xinyi Chen, Evan Dogariu, Devan Shah, Hubert Strauss, Vladimir Feinberg, Daniel Suo, Peter Bartlett, Elad Hazan ICLR 2026 website / paper / code / slides |
|
ScooterId: Posture-Based Continuous User Identification From Mobility Scooter RidesDevan Shah , Ruoqi Huang, Nisha Vinayaga-Sureshkanth, Tingting Chen, Murtuza Jadliwala IEEE TMC 2025 paper |
Projects |
|
Scaling Mutual Information Skill Learning to Larger Skill SpacesDevan Shah*, David Shustin* (* indicates equal contribution) paper |
Surveys |
|
Discussion on "Why Does Deep and Cheap Learning Work So Well?"Devan Shah paper |
|
A Survey of State Space Models: From Linear Systems to LanguageDevan Shah* , Brandon Cho* paper |
|
Understanding Dynamic Algorithms for Packing-Covering LPs via Multiplicative Weight UpdatesDevan Shah* , Owen Yang*, Sunil Vittal* paper |
Coursework |
Computer Science |
||||||||||||||||||||
|
||||||||||||||||||||
Math |
||||||||||||||||||||
|
Theory |
||||||||||||||||
|
||||||||||||||||
Creative Writing |
||||||||||||||||
|
Contests |
|