A picture of Aditya

Aditya Parashar

aparashar@umass.edu // CV // LinkedIn // SCHOLAR // twitter

I am a Masters student in Computer Science at UMass Amherst with a strong focus on natural language processing and LLM research. My research focuses on enhancing the reasoning abilities of large language models (LLMs) and improving their inference through probabilistic sampling and decoding strategies, while also exploring their applications in specialized fields like operations research and healthcare. I have collaborated with leading researchers at Oracle, Google, IESL, and the BioNLP lab at UMass.

Before starting my masters, I worked at LTRC, IIIT Hyderabad in the area of machine translation for Indian languages. Prior to that, I built distributed software applications at Agility and even before that, I spent four wonderful years at IIT Guwahati earning my bachelors in mechanical engineering and mathematics & computing.

I am looking for full time opportunities starting January 2025, in machine learning and NLP, please reach out if you think I would be a good fit.

Projects

Leveraging LLMs for what-if analysis of Operations Research problems
Summer 2024
Developed a dataset to test LLMs on operations research problem reformulation through what-if questions, revealing their struggles with increased problem complexity. Work done during summer internship at Oracle labs with Natalia Kosilova and Ari Kobren
Quasi-random Multi-Sample Inference for Large Language Models
Spring 2024
As part of an industrial independent study with Luke Vilnis, implemented arithmetic sampling for mult-sample inference from LLMs. Evaluated it for generating diverse candidates for chain-of-thought self-consistency for reasoning problems and for minimum bayes risk decoding for machine translation.
Multiple Choice Question Generation and Evaluation With Critique, Correction, and Comparison Feedback
Summer 2023
Implemented a question generation-question answering system using a self-refining LLM(GPT-3+) based framework, working with Zonghai Yao under the guidance of Prof. Hong Yu.
Disentangling LLM probability normalization methods
Fall 2023
Working with Dhruv Agarwal on various probability calibration techniques to mitigate surface form competition, as part of an Independent study under Prof. Andrew Mccallum.
Spill the Implicit Beans: Interpreting figurative language using prompting
Spring 2023
Developed a framework to directly translate figurative language into literal form using reasoning-based prompting strategies, significantly improving human evaluation scores across metaphors, idioms, similes, and sarcasm.
Interlingua(USR) based machine translation
2022
Worked on a deep learning+heuristics based machine translation strategy for Indian languages with Dr. Sukhada and Dr. Soma Paul

This website is built based on Prof. Mohit Iyyer's website template.