Chinmaya Kausik

• Tackling tangible, real-world questions with a principled mathematical approach. These days, my PhD research focuses on sequential decision making under various settings - offline-to-online transfer, partial observability/latent information and non-standard feedback and reward models. I also have side projects and internship research in deep learning, LLM agents, transformers, Bayesian inference, etc. On the other hand, a lot of my undergraduate background was in geometry, topology and dynamics, with work in computer-assisted topology and geometry.
• Increasing accessibility to and in higher mathematics and creating communities where ideas cross pollinate and people pull each other up. I have started the Stats, Physics, Astronomy, Math (SPAM) graduate student social initiative at the University of Michigan. I also co-founded and co-organize Monsoon Math Camp. I have also been involved in building and expanding other mathematical communities, like platforms for the PolyMath REU, DRP programs and the undergraduate math organization at IISc, etc.

• Working on how we can robustly combine offline models and online interaction, in contrast to combining offline data and online interaction. With pretrained models like LLMs, foundation models for protein folding, etc readily available but lacking guarantees, we need principled sequential decision making algorithms that use offline models that don’t have performance guarantees.
• Continuing the LLM agent research from my Nteflix internship, where we designed a generic retrieval/memory agent and we are not formalizing the problem to design more informed experiments.

• Writing a paper based on my internship at Microsoft in the advertiser optimization team under Ajith Moparthi! I designed and implemented a fast algorithm for updating models used for advertiser bidding.

• Organizing an interdepartmental social initiative, SPAM (Statistics, Physics, Astronomy, Mathematics).
• Fleshing out ideas for more academic communities like Monsoon Math.

primary goals

• Conduct experiments on the efficacy of sequence model architectures with explicit separate latent states (TRMs, HRMs, the free transformer), and conduct an extensive interpretability study of their latent states.

• Work on designing an agent/model to automatically annotate graded UMich calculus answer sheets, focusing on preventing hallucinations.

• Work on a large scale applied recommender systems project using the latent bandit algorithms that I designed (LOCAL-UCB and ProBALL-UCB).

side-quests

• Design codenames bots at both the agent level and by training them from scratch, both with separate as well as combined spymasters and guessers.
• Explore the nuances of implementing various RL algorithms in simulated motion settings.