Saraswathy Amjith
AI Researcher
Research Overview
Enhancing Large Language Models Through Self-Play and Reinforcement Learning
My research at MIT's Computer Science & Artificial Intelligence Laboratory focuses on developing novel approaches to improve large language model capabilities through game-theoretic frameworks and reinforcement learning. I'm particularly interested in creating self-improving systems where AI agents compete and learn from each other, pushing the boundaries of reasoning and problem-solving in artificial intelligence.
In-Context Learning for Esoteric Programming Languages
Developed a novel in-context reinforcement learning approach that enhances LLM performance on highly specialized programming tasks without the computational cost of fine-tuning or retraining.
Research Question: Can we enable LLMs to rapidly adapt to domain-specific languages through contextual learning alone?
Impact: Demonstrates efficient knowledge transfer in constrained computational environments.
Read Paper at NeurIPS 2025 →
Click to view full research poster presented at NeurIPS 2025
A Novel Integrated ML Approach Utilizing Radar & Satellite Imagery for Detecting Illegal Selective Logging
Innovated AI models for detecting deforestation by integrating radar and satellite data (a first in machine learning) to counteract the challenges posed by selective logging and cloud cover interference, innovating CNNs, Random Forests, and XGBoost to achieving a 7% improvement in detection accuracy over previous studies, reaching a 95.08% accuracy rate.
Engineered a data fusion pipeline that merged over 134,000 geospatially aligned satellite and radar images into unified training samples by synchronizing temporal frames, resulting in the largest dataset to date for detecting illegal deforestation.
Publications
About
I'm an undergraduate student at MIT pursuing Computer Science. As an undergraduate researcher at MIT's Computer Science & Artificial Intelligence Laboratory (CSAIL), I work on advancing large language model capabilities through novel reinforcement learning approaches.
My research centers on understanding how LLMs learn to reason and improving their abilities to do so. I'm also interested in alignment.
Research Interests: Large Language Models, Reinforcement Learning, Game Theory, In-Context Learning, AI Safety