Things I've Worked On 🍜

This paper reproduces and stress-tests MEMIT, a method for editing factual knowledge into LLM weights, finding it works perfectly on GPT-J but transfers poorly to Llama 3.1 8B. Even when edits succeed, clever adversarial prompts can trick the model into recalling its original, pre-edit knowledge over 65% of the time.

Developed a novel data structure for efficiently analyzing product metrics by extending classical range tree techniques. Designed and implemented greedy range trees in Python to optimize multi-dimensional queries, improving speed and scalability for large datasets. This enables faster, more accurate analysis of product metrics, including trends, anomalies, and usage patterns.

A compilation of selected problems and solutions from research in advanced graph theory, covering topics such as graph colorings, connectivity, planarity, and extremal problems.

In 1928, Emil Artin and Helmut Hasse introduced the Artin–Hasse exponential, a p-adic analogue. After introducing the basics of p-adic analysis, including why \(\mathbb{Q}_p\) is the completion of \(\mathbb{Q}\), this paper shows that the exponential and logarithm remain inverses, with intuition drawn from metric spaces and topology. The core is an inductive-based proof of Dwork's Lemma, used to prove the integrality of the Artin–Hasse Exponential.

Building on Christiaan Huygens' discovery of coupled pendulum synchronization, this paper examines how string length affects synchronization time on a moving platform. Experiments showed that shorter strings generally synchronize faster. Behaviors like brief pauses and in-phase vs. anti-phase motion are explained using classical mechanics. The paper also models motion and energy using Lagrangian techniques.