Alok Shah

I'm a student at the University of Pennsylvania, where I study Computer Science, Mathematics, and Electrical Engineering.

Outside of class, I lead MLR@Penn, the undergraduate AI and ML research organization and community at Penn and serve on the Student Advisory Board for the Wharton AI & Analytics Initiative.

In my free time, I enjoy historical documentaries, geography trivia, trying new food trucks and restaurants around Philly, and burning off those calories by playing pickup soccer.

Email / CV / Twitter / Github

Projects

I play simple tricks on machine learning models to study their training dynamics. I tend to interpret the phenomena which emerge from these experiments with a lens informed by theory. I'm interested in applying such insights to problems in robotics and language.

	Dropout, Johnson-Lindenstrauss, Low-Rank Bias, and Generalization Alok Shah, Mohul Aggarwal, Khush Gupta CIS 6770 Final Project: An alternate perspective on the generalizing behavior of dropout through the lens of random projections and low-rank bias
	Language Modeling With Learned Meta-Tokens Alok Shah, Khush Gupta, Keshav Ramji, Pratik Chaudhari ICML LCFM, 2025 Showed that meta-tokens provably improve context propagation in language models across extreme lengths via rate-distortion analysis and positional ablations
	Investigating Language Model Dynamics using Meta-Tokens Alok Shah, Khush Gupta, Keshav Ramji, Vedant Gaur NeurIPS ATTRIB, 2024 ESE 5460 Final Project: Explored how to coerce communication between token-level checkpoints for more interpretable, capable models
	Modeling Human Behavior Without Humans: Prospect Theoretic Multi-Agent RL Sheyan Lalmohammed, Khush Gupta, Alok Shah, Keshav Ramji, ICML MAS*, 2025 STAT 4830 Final Project: A cool way to induce human-like risk-averse behavior between multiple agents by warping the reward landscape in Choquet integrals in accordance with Cumulative Prospect Theory
	Weak-to-Strong In-Context Optimization of Language Model Reasoning Keshav Ramji, Alok Shah, Vedant Gaur, Khush Gupta NeurIPS ATTRIB, 2024 Developed in-context optimization method leveraging weak learners to improve reasoning in strong large language models without additional finetuning
	A Note on Multitask Learning with Task-Adaptive Priors Alok Shah , Sidhant Srivastava, Tara Kapoor CIS 7000 Final Project: By extracting task dependent statistics, we learn a prior that provably nudges the model toward representations that generalize better to new tasks.
	Conformal Actor-Critic: Distriution-Free Uncertainty Quantification for Offline RL Alok Shah, Nikhil Kumar, Khush Gupta, Mohul Aggarwal CIS 6200 Final Project: Integrated conformal prediction into offline reinforcement learning, providing statistically robust uncertainty quantification to curb overestimation bias of Q-values
	Deep Compression with Adversarial Robustness via Decision Boundary Smoothing Alok Shah, Michael Shao ESE 5390 Final Project: Smoothed adversarial retraining during compression produces compact models with stronger robustness
	AlfLLM: Limitations on LLMs as Reward Function Surrogates Alexander Kyimpopkin, Alok Shah, Dominic Olaguera-Delogu ESE 6500 Final Project: Investigated the efficacy of using large language models as surrogate reward functions for reinforcement learning in the ALFWorld environment.