About
I’m a master’s student in Artificial Intelligence at the Gwangju Institute of Science and Technology (GIST), and a graduate research assistant at the Data Science Lab with Prof. Sundong Kim. I work on ARC-AGI-3, building agentic systems that combine reinforcement learning and large language models for stronger reasoning and generalization, targeting state-of-the-art performance on the benchmark.
ARC-AGI-3
Reinforcement Learning
Large Language Models
Education
2025 — now
M.S. in Artificial Intelligence
Gwangju Institute of Science and Technology (GIST)
2019 — 2024
B.Sc. in Computer Engineering
University of Tehran
Research
2026
ARC-AGI-3: learning & planning in unknown environments
Data Science Lab · GIST
Agentic systems combining RL, search and memory for interactive games with no external supervisor; LLM-synthesized world models for planning. More ↓
ARC-AGI-3 is a benchmark of interactive games where an agent must infer mechanics and goals purely through interaction, under offline evaluation constraints with no external supervisor. My main focus is developing agentic systems that combine reinforcement learning, search, and case-based memory for efficient decision-making and generalization, targeting state-of-the-art performance. I am also investigating the use of large language models to synthesize executable world models from interaction traces to support downstream planning.
2024
Transfer learning in RL using LLMs
Cognitive Systems Lab · University of Tehran
Bachelor’s thesis: LLMs identify similarities between past knowledge and new tasks to accelerate RL. More ↓
Reinforcement learning often faces long training times and difficulty adapting to new problems. My thesis applied transfer learning to carry knowledge from past experience into new tasks, using large language models to identify similarities and differences between prior knowledge and the new problem. Experiments showed that LLMs can significantly accelerate learning and improve decision-making compared to traditional RL techniques. Poster →
2023
Educational technology
Cognitive Systems Lab · University of Tehran
Adaptive learning with RL, LLM fine-tuning & RLHF for educational content, and a custom LMS. More ↓
We researched reinforcement learning for tailored adaptive learning systems and used LLMs — via fine-tuning and RLHF — to produce bespoke educational content. We also introduced a novel pedagogy model for deeper learning and built a learning management system with Python, Django, Bootstrap and JavaScript, integrating an optimized ChatGPT with Google Classroom alongside tools like Discord and AI-based robots. Website →
2022
Hate speech detection
Intelligent Information Systems Lab · University of Tehran
Jointly trained GCN + BERT over a heterogeneous multilingual graph of words, tweets and hashtags. More ↓
We jointly trained GCN (Graph Convolutional Networks) and BERT to detect hate speech. Documents are represented as nodes in a heterogeneous multilayer graph of words, tweets and hashtags, with inter-layer edges denoting semantic equivalence across languages — allowing local and global information to interact in a comprehensive classification representation.
2021
Crystalline
Data Analytics Lab · University of Tehran
Proof-of-concept cryptocurrency in pure Python with a novel proof-of-activity consensus. More ↓
We developed a proof-of-concept cryptocurrency in pure Python, introducing a unique proof of activity (PoA) consensus protocol crafted to combine the advantages of both PoS and PoA. Repo →
Publications
A Non-LLM Graph-Search Agent for ARC-AGI-3 with Novelty-Guided Exploration and Case-Based Memory
Mohammad Saadati, Laura Mascarell, Sundong Kim
Mohammad Saadati, Laura Mascarell, Sundong Kim
RPoA: Redefined Proof of Activity
Sina Kamali, Shayan Shabihi, Mohammad Taha Fakharian, Alireza Arbabi, Pouriya Tajmehrabi, Mohammad Saadati, Behnam Bahrak
Sina Kamali, Shayan Shabihi, Mohammad Taha Fakharian, Alireza Arbabi, Pouriya Tajmehrabi, Mohammad Saadati, Behnam Bahrak
Off the clock
Lifelong football fan — Hala Madrid ⚽ — and an avid FPS gamer, usually somewhere on a Battlefield server.