Prakhar Dixit

PhD Student, Computer Science
University of Maryland, Baltimore County

I am a PhD student in Computer Science at the University of Maryland, Baltimore County (UMBC). My research asks how AI systems reason, learn from experience, and use memory.

Most of my current work is on mathematical reasoning in large language models and on memory for LLM agents: how an agent can accumulate reusable problem-solving strategies over time (ISM), and why agents sometimes confidently misremember their own past (Honest Lying). Earlier, I worked on sample efficiency in reinforcement learning and on sim-to-real transfer for vision-based drone navigation.

In summer 2026 I was an Applied Scientist Intern at Microsoft, working on agent skills in Copilot Studio. Before that I was a research intern at Emergence AI, working on long-term memory for conversational agents.

Email CV Google Scholar GitHub LinkedIn

Research Interests

• Mathematical reasoning in LLMs — strategy memory, schema-based reasoning, learning from feedback
• Memory for LLM agents — long-term memory, continual learning, confabulation and reliability
• Reinforcement learning — sample efficiency, hierarchical agents, graph-based representations, sim-to-real transfer

News

• Aug 2026Completed my Applied Scientist internship at Microsoft (Copilot Studio).
• Jun 2026ISM accepted at the AI for Math Workshop at ICML 2026.
• Jun 2026Honest Lying accepted at the Workshop on Failure Modes in Agentic AI at ICML 2026.
• May 2026Started as an Applied Scientist Intern at Microsoft (Copilot Studio), Redmond.
• Mar 2026WildfireVLM accepted at IGARSS 2026.
• Jul 2025Our long-term memory system at Emergence AI reached 86% on LongMemEval (blog post).
• May 2025Joined Emergence AI in New York as a summer research intern.
• Nov 2024SBI-RAG accepted at the MATH-AI Workshop at NeurIPS 2024.
• Aug 2023Started my PhD in Computer Science at UMBC.

Selected Publications

• ICML 2026
  Workshop
  ISM: Self-Improving Strategy Memory for Continual Mathematical Reasoning

  Prakhar Dixit and Tim Oates

  AI for Math Workshop (AI4Math) at ICML 2026

  A strategy memory that learns from both successful and failed episodes to improve mathematical reasoning with a frozen language model, evaluated on MATH-Hard and OlympiadBench.

  arXivOpenReviewCode
• ICML 2026
  Workshop
  Honest Lying: Understanding Memory Confabulation in Reflexive Agents

  Prakhar Dixit, Sadia Kamal, and Tim Oates

  Workshop on Failure Modes in Agentic AI at ICML 2026

  Studies why reflexive agents produce plausible but incorrect recollections of their own past, and what this implies for the reliability of long-horizon agents.

  arXiv
• NeurIPS 2024
  Workshop
  SBI-RAG: Enhancing Math Word Problem Solving for Students through Schema-Based Instruction and Retrieval-Augmented Generation

  Prakhar Dixit and Tim Oates

  MATH-AI Workshop at NeurIPS 2024

  Combines schema-based instruction with retrieval-augmented generation to produce step-by-step, schema-guided solutions to math word problems.

  arXiv
• AAAI 2023
  Workshop
  ReProHRL: Towards Multi-Goal Navigation in the Real World using Hierarchical Agents

  Tejaswini Manjunath, Mozhgan Navardi, Prakhar Dixit, Bharat Prakash, and Tinoosh Mohsenin

  Reinforcement Learning Ready for Production Workshop at AAAI 2023

  Hierarchical RL for multi-goal navigation, transferred from simulation to a real nano-drone.

  arXiv

Full publication list →

Academic Service

• Reviewer, 6th Workshop on Mathematical Reasoning and AI (MATH-AI) at NeurIPS 2026
• Reviewer, AI for Math Workshop at ICML 2026
• Reviewer, MATH-AI Workshop at NeurIPS 2025
• Reviewer, ACL 2025 (3 papers)
• Reviewer, Inductive Biases in Reinforcement Learning Workshop at RLC 2025
• Talks and tutorials reviewer, PyCon US 2025