ISM: Self-Improving Strategy Memory for Continual Mathematical Reasoning
A strategy memory that learns from successful and failed episodes to improve mathematical reasoning with a frozen language model. Evaluated on MATH-Hard and OlympiadBench.
Computer Science · UMBC
Prakhar Dixit
Mathematical reasoning. Agent memory. Reinforcement learning.
I am a PhD student in Computer Science at the University of Maryland, Baltimore County. I study how AI systems reason, learn from experience, and use memory, from mathematical problem solving with language models to reinforcement learning and robotics.
Featured Research
Research Interests
LLMs & Mathematical Reasoning
Schema-based reasoning, self-improving strategy memory, and learning from feedback in language models.
Reinforcement Learning
Deep RL, hierarchical agents, Sim2Real transfer, graph attention in R-GCN frameworks, sample efficiency.
Computer Vision & Robotics
YOLO-based object detection, autonomous drone navigation, real-world RL deployment.
Recent Highlights
Jun 2026ISM: Self-Improving Strategy Memory for Continual Mathematical Reasoning, accepted to the AI4Math Workshop at ICML 2026.
Jun 2026
Paper accepted at
ICML 2026 · Failure Modes in Agentic AI Workshop
— Honest Lying: Understanding Memory Confabulation in Reflexive Agents.
May 2026
Started as
Applied Scientist Intern at Microsoft
— working with the Copilot Studio Team to improve skills inside the MCS platform.
Mar 2026
Paper accepted at
IGARSS 2026
— WildfireVLM: AI-powered Early Wildfire Detection and Risk Assessment Using Satellite Imagery.
Jul 2025
Published SOTA results on
LongMemEval
benchmark (86% accuracy, +15% over previous SOTA) at Emergence AI.
Nov 2024
Paper accepted at
NeurIPS 2024 · MATH-AI Workshop
— SBI-RAG: Enhancing Math Word Problem Solving via Schema-Based Instruction and RAG.
Aug 2024
MS thesis published on ProQuest — Dynamic Edge Weighting in R-GCN for Sample Efficiency in RL.
Jan 2023
Paper accepted at
AAAI 2023 · RL Ready for Production Workshop
— ReProHRL: Multi-Goal Navigation using Hierarchical Agents.
Featured Projects
ScholarAgent
An AI agent that fetches recent ArXiv research papers based on search — like Google Scholar but more user-friendly and readable.
LLM Detect AI Text
ML model to detect whether an essay was written by a student or an LLM. Finished in the top 25% of the Kaggle leaderboard.
ISS_Info
Python package fetching real-time International Space Station data. Over 60,000 downloads on PyPI.
Deep Learning Steering Model
Self-driving prototype using CNNs and OpenCV to predict steering angles from dash-cam image inputs.