// about

Teaching language models to reason, and checking the work.

I am a Ph.D. candidate in Computer Science and Engineering at Pennsylvania State University, advised by Prof. Vasant Honavar, graduating in November 2026. My research is on reasoning in large language models. I build agentic workflows and evaluation harnesses that find where models fail, generate targeted synthetic data to fix those failures, and feed it back through SFT, RL, and pretraining.

Most recently I was a Research Scientist Intern at a stealth AI startup in New York, training reward models for creative judgment. Before that I collaborated with Google's Gemini Data team on the Gemini IMO 2025 project, and my work on LLMs for mathematics is supported by an unrestricted research gift from Google.

// news

News

Oct 2026
Splicing Hints into Reasoning Traces is on arXiv and under review at ICLR 2027.
Aug 2026
Wrapped up a six-month Research Scientist internship at a stealth AI startup in New York.
Mar 2026
Joined a stealth AI startup in New York to work on reward models for creative judgment.
Oct 2025
RefGrader and CombiGraph-Vis accepted to the Math-AI Workshop at NeurIPS 2025.
Jul 2025
Collaborated with Google's Gemini Data team on the IMO 2025 gold medal effort.
May 2025
Received a $100k unrestricted research gift from Google.
Apr 2025
Brains vs. Bytes accepted to COLM 2025.
// publications

Publications

Submitted to ICLR2027
Splicing Hints into Reasoning Traces: Distillation Beyond What the Demonstrator Can Solve
H. Mahdavi, P. Mahdavinia, M. Farhadi, V. Honavar
arXiv
Math-AI Workshop, NeurIPS2025
RefGrader: Automated Grading of Mathematical Competition Proofs using Agentic Workflows
H. Mahdavi, A. Hashemi, M. Daliri, P. Mohammadipour, A. Farhadi, S. Malek, Y. Yazdanifard, N. Mireshghallah, A. Khasahmadi, V. Honavar
arXiv
Math-AI Workshop, NeurIPS2025
CombiGraph-Vis: A Multimodal Olympiad Benchmark for Discrete Mathematical Reasoning
H. Mahdavi, A. Hashemi, M. Daliri, P. Mohammadipour, A. Farhadi, S. Malek, Y. Yazdanifard, A. Khasahmadi, V. Honavar
arXivDataset
Preprint2025
Harnessing Optimization Dynamics for Curvature-Informed Model Merging
P. Mahdavinia, H. Mahdavi, N. Mireshghallah, M. Mahdavi
arXiv
COLM2025
Brains vs. Bytes: Evaluating LLM Proficiency in Olympiad Mathematics
H. Mahdavi, A. Hashemi, M. Daliri, P. Mohammadipour, A. Farhadi, S. Malek, Y. Yazdanifard, A. Khasahmadi, V. Honavar
arXiv
Digital Discovery2025
Beyond Training Data: How Elemental Features Enhance ML-Based Formation Energy Predictions
H. Mahdavi, V. Honavar, D. Morgan
Paper
AAMAS2020
Towards Deployment of Robust Cooperative AI Agents: An Algorithmic Framework for Learning Adaptive Policies
A. Ghosh, S. Tschiatschek, H. Mahdavi, A. Singla
arXiv
// experience

Experience

Research Scientist Intern
Mar 2026 – Aug 2026
Stealth AI Startup · New York, NY
  • Developed agentic workflows and evaluation harnesses for open-ended creative content generation, with a focus on generating synthetic reasoning data for creativity.
  • Built pipelines that transform agentic workflows into training data, studying the interaction between mid-training and post-training for creative reasoning capabilities.
  • Trained reasoning and non-reasoning reward models for creative judgment and led data curation for reward modeling in the creative judgment domain.
Collaborator, Gemini IMO 2025 Project
Summer 2025
Google · New York, NY
  • Provided high-quality verifiable data and evaluations to the Gemini Data team under the supervision of Dr. Vahab Mirrokni, resulting in a gold medal at the International Mathematical Olympiad (IMO).
Research Assistant
Jan 2022 – Present
Pennsylvania State University · State College, PA
  • Design agentic workflows for math problem solving and high-quality mathematical data generation; build and run evaluation harnesses for contest-level math (benchmarks, metrics, ablations, failure analysis).
Data Scientist Intern
Jun 2023 – Aug 2023
Arconic Technology Center · New Kensington, PA
  • Worked in the Fabrication Technology group under the supervision of Martin Marinack; tackled several data-driven optimization problems to improve Arconic's fabrication processes (welding, rolling, and casting).
Research Intern
Aug 2018 – Jan 2019
Max Planck Institute for Software Systems · Saarbrücken, Germany
  • Worked on fundamental aspects of reinforcement learning, focusing on multi-agent, adversarial, and human-centered reinforcement learning.
  • Studied how to design an AI agent that can robustly cooperate with agents of unknown types encountered only after deployment.
Software Engineering Intern
Jul 2017 – Sep 2017
Faragostar · Tehran
  • Developed a compression algorithm that allowed Faragostar to retain its official documents while minimizing storage requirements.
  • Configured infrastructure for Elasticsearch, enabling more efficient and complex searches than the previous SQL-based search tool.
Mathematical Olympiad Teacher
Oct 2012 – Apr 2013
Allameh Helli High School · Tehran
  • Taught number theory to academically gifted students.
// education

Education

Ph.D. in Computer Science and Engineering
Pennsylvania State University · Jan 2022 – Nov 2026 (expected)
Advisor: Prof. Vasant Honavar · Thesis: Solving Contest-Level Math Problems Using LLMs
M.Sc. in Artificial Intelligence
Sharif University of Technology · Sep 2017 – Dec 2020
Advisor: Prof. Hamid Beygi · Thesis: Approximate Inference in Probabilistic Network Models
B.Sc. in Computer Engineering
Sharif University of Technology · Sep 2012 – Sep 2017
Advisor: Prof. Hossein Sameti · Thesis: A Survey on Speech Enhancement Algorithms
// honors

Honors

$100k Unrestricted Research Gift from Google
May 2025 · Awarded to support research on enhancing LLMs for mathematics through synthetic data generation and agentic workflows.
Top National Rankings in M.Sc. Entrance Exam
Jun 2017 · Placed 5th (AI), 2nd (Networks and Security), and 4th (IT) nationwide.
Ranked 217th in National University Entrance Exam
Jun 2012 · Top 0.1% among 260,000+ examinees nationwide.
Silver Medal, 29th National Mathematical Olympiad
Sep 2011 · Selected as one of 40 participants from 100,000+ candidates.
// interests

Interests

  • LLMs for math problem solving
  • Agentic workflows for synthetic data generation
  • Reasoning in LLMs
  • Post-training, reward modeling, and LLM-as-a-judge
  • Reinforcement learning
  • Optimization in machine learning

Outside research I play tennis and badminton, which I have done since I was nine, and guitar, acoustic and electric, mostly classic rock.