Masoud Jafaripour

Masoud Jafaripour

M.Sc. Student Researcher
University of Alberta



Bio

I am an M.Sc. student researcher in the Department of Computing Science at the University of Alberta, supervised by Prof. Osmar Zaiane and collaborating with Prof. Matt Taylor. I am also a Research Scientist Intern at Electronic Arts (EA), training small language models for long-horizon, multi-turn agentic tasks in real-time gameplay. My thesis studies visual and spatial reasoning in multimodal language models. I work on adaptive, efficient 3D spatial reasoning and action prediction for multimodal agents, including inference-time scaling and context compression, on curating vision-language datasets for embodied agents, and on improving multimodal reasoning through in-context learning, SFT, and RL fine-tuning.

More about me

Before moving to Computing Science, I was a Ph.D. student in Electrical & Computer Engineering at the University of Alberta, where I worked on robotics. I have 3+ years of academic and industry experience building and deploying deep learning systems. I have also worked on multimodal retrieval (visual RAG, geometric retrieval), visual-tabular representation learning, 3D CAD generation, uncertainty quantification in LLMs, and deep RL for continuous control.

News

[Now] I am a Research Scientist Intern at Electronic Arts (EA) in Vancouver, training small language models for long-horizon agentic tasks in real-time gameplay.
[Oct 26] Our preprint "Towards a Data-Centric Understanding of Spatial Reasoning in Multimodal Large Language Models" is on arXiv and under review at Artificial Intelligence Review.
[2025] Released Vibe-CADing, an open-source text-to-CAD pipeline where an LLM writes and repairs CAD programs.
[2025] Our survey "The Evolving Landscape of LLM- and VLM-Integrated Reinforcement Learning" is accepted to IJCAI 2025.
[2025] Our paper "Adaptive Iterative Feedback Prompting for Obstacle-Aware Path Planning via LLMs" is accepted to the LM4Plan Workshop @ AAAI 2025.
[Feb 25] Our survey on LLM- and VLM-integrated RL is on arXiv.
[Oct 24] Our paper "Optimal Integration of Hybrid FES-Exoskeleton for Precise Knee Trajectory Control" appears at IROS 2024.

Research Experience

Electronic Arts (EA) (Research Scientist Intern)Post-training small language models for long-horizon, multi-turn agentic gameplay as action & world models 05/26 – Now Computing Science, University of Alberta (Research Assistant)Implicit world modeling and spatial-action reasoning in VLAs, inference-time scaling in multimodal LMs 06/24 – Now Vector Institute (Teaching Assistant)Multimodal LLM Bootcamp, guiding teams building multimodal pipelines 08/25 – 12/25 Amii (Alberta Machine Intelligence Institute) (Work Integrated Learning)Multimodal pipelines for real/fake image detection, AI ethics workshops 08/25 – 03/26 Stanford NLP Group (NLP Researcher, remote)LLM uncertainty estimation via contrastive prompting and graph-based confidence mapping 01/25 – 05/25 iSMART & ECE, University of Alberta (Research Assistant)Real-time torque optimization and deep RL control for robotic exoskeletons 09/22 – 06/24

Publications

  1. SAVeR: Spatially-Adaptive Visual Reasoning with Self-Reward in Multimodal LLMs
    Masoud Jafaripour, Mohamadali Tofigh, Osmar R. Zaiane
    Under review
  2. Towards a Data-Centric Understanding of Spatial Reasoning in Multimodal Large Language Models
    Mahdieh Mallahnezhad, Masoud Jafaripour, Mohamadali Tofigh, Sheila Schoepp, Osmar R. Zaiane
    Under review at Artificial Intelligence Review PDF
  3. The Evolving Landscape of LLM- and VLM-Integrated Reinforcement Learning
    Sheila Schoepp, Masoud Jafaripour*, Yingyue Cao*, Tianpei Yang, Fatemeh Abdollahi, Shadan Golestan, Zahin Sufiyan, Osmar R. Zaiane, Matthew E. Taylor (*Equal contribution)
    International Joint Conference on Artificial Intelligence (IJCAI), 2025, Montreal, Canada PDF Code 19.6% acceptance rate
  4. Adaptive Iterative Feedback Prompting for Obstacle-Aware Path Planning via LLMs
    Masoud Jafaripour, Shadan Golestan, Shotaro Miwa, Yoshihiro Mitsuka, Osmar R. Zaiane
    LM4Plan Workshop @ AAAI Conference on Artificial Intelligence (AAAI), 2025, Philadelphia, USA PDF Code
  5. Optimal Integration of Hybrid FES-Exoskeleton for Precise Knee Trajectory Control
    Masoud Jafaripour, Vivian K. Mushahwar, Mahdi Tavakoli
    IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, Abu Dhabi, UAE PDF

Selected Projects

nanochat-VLM Oct. 2025 – Ongoing
RustBPE/tiktoken, FineWeb, PyTorch DDP · Personal
A minimal VLM built from scratch by extending nanochat with vision tokenization and multimodal fusion.
Code
Vibe-CADing Featured Apr. 2025 – Ongoing
Diffusion, AR Generation, LangChain, B-Rep, Harness, Blender · Personal
An agentic harness pipeline that retrieves, generates, and refines CAD parts from natural-language prompts and visual GUI input.
Project Page Code Report
Visual RAG Feb. 2025 – Jun. 2025
RAG, FAISS, SAM, DINO, CLIP · Amii Labs
A hybrid image-text retrieval system that fetches relevant demonstrations to improve few-shot in-context planning.
Spatial Reasoning in VLMs Jan. 2025 – May 2025
LLaVA/Qwen-VL, Transformers & TRL · Reasoning Course
Enhanced spatial reasoning in VLMs with CoT and VoT prompting and in-context RL for 3D visual environments.
Code
Game Pathfinding Feb. 2025 – Apr. 2025
C++, Docker, Heuristic Search, Graph Data Structures · Search Course
Efficient search through offline abstraction and online heuristic embeddings for pathfinding on game maps.
Code Report

Powered by Jekyll and Minimal Light theme.