Maxwell W. Barlow

DW22-284

11 Gore Rd.

Webster, MA 01570

Introducion

I am an aspiring Research Scientist searching for a lab with which to complete an assistantship and to pursure a PhD beginning in Fall ‘25. I am currently working as a Software Engineer for MAPFRE Insurance, modifying and configuring the insurance policy management software Guidewire using Gosu (a JVM language). My main research interests lie in Natural Language Processing, Ethical AI, and model interpretability.

Research

I have done extensive work with pre-trained Transformers such as OpenAI’s GPT-3.5, Meta’s OPT, and Google’s T5, investigating vector representations of tokens called “embeddings” for applications in transfer learning, fine tuning, and interdisciplinary investigation. Moving forward, I am interested programmatically investigating and characterizing “truth” in embeddings and other artifacts of artificial language generation to further our understandings of how Large Language Models generate output and can come to “hallucinate”. Principally, I would be investigating the direction (or lack thereof) of truth in embeddings from Large Language Models. On a broader scale, I am determined to investigate how our efforts towards AI alignment (and fine tuning in general) modifies the hyperdimensional “plane of best fit” modeled by the transformer architecture. For example, I am curious how the same vector embedding may change in use within a model while that model undergoes different kinds of fine-tuning. Particularly when we fine-tune to remove direct utterances of predjudiced ideas, do algorithms simply trim problematic outputs, redistribute explicit harm into other less explicit outputs, or something else entirely? And how does the model itself reallocate the space previously used to hold predjudiced language and structures?

Other Interests

I have spent past lives as a chamber choir singer, as well as a songwriter and guitarist. I got my start as a teen with District and State choirs in Massachusetts, and kindled a love for live performance at open mic’s with a ukulele in the mid 2010’s. I love live music and the occasional craft beer, so catch me at a local brewery for community events and local bands! Music is a continued interest of mine, and I hope to bring it with me wherever I go.

selected publications

MOCSIDE: An Open-Source and Scalable Online IDE and Auto-Grader for Computer Science Education

Jonathan Cazalas, Max Barlow, Ibraheem Cazalas, and Chase Robinson

In Proceedings of the 53rd ACM Technical Symposium on Computer Science Education V. 2, 2022

Abs

Programming is learned through practice, with said practice in introductory programming courses often translating to a prohibitively large number of assignments, increasing the grading workload for faculty and/or teaching assistants. In short, this is unsustainable. Several publishers and a few notable companies have provided meritable auto-grading solutions, although most are plagued with problems including minimal problem sets, limited customization options, high cost, and at times even a disconnect with the pedagogical needs of academia. This poster presents our newly-developed web application, MOCSIDE, an open-source and scalable online IDE and auto-grader for computer science education. Results indicate a positive user experience from students and instructors alike, with cost savings, ease of use, and code collaboration highlighted as key features.
Going with the Flow: An AI Approach to Solving the Puzzle Game FlowFree

Maxwell Barlow, Spencer Riekens, William Bennett, and Matthew Eicholtz

In Proceedings of the 2023 Florida Undergraduate Research Conference, Feb 2023

Abs

Can we solve FlowFree? FlowFree, also known as Zig-Zag Numberlink, is a popular puzzle game where the player connects ‘flows’ on a board with the goal of finishing each connection, avoiding intersecting flows, and filling up the entire game board. We employed two methods for solving puzzles: A* search and boolean satisfiability (SAT). A* searched through a tree of board states, choosing between states one cell apart based on a heuristic that considered distance to a completed flow. The search tree was pruned with the help of several constraints of varying complexity to check for. Constraint extrapolation further allowed the framing of FlowFree as a constraint satisfaction problem (CSP), which allowed the solving of boards as SAT problems. Solving as a CSP involves the conversion of puzzles into SAT variables, indicating the truth value of possible flow directions and color at each individual cell, arranged into conjunctive normal form clauses representing the constraints, feeding to a SAT solver and subsequently decoded from a set of true variables to a solved puzzle. SAT showed significant speed improvement over A*, with the difference being more pronounced in larger puzzle sizes. Puzzles were collected from the FlowFree mobile app and converted using computer vision to text files; A GUI application was also created to allow inputting custom puzzles as well as partially solving a puzzle before solve time.