Learning via Self-Consistency for Diffusion-based Video Reasoning
arXiv preprint, 2026
About Me
Hi, I'm Zhenghao Ni, a Computer Engineering undergraduate at the University of Toronto. My work spans diffusion-based video reasoning, world models, and efficient machine learning systems.
I am currently a Machine Learning Engineer Intern at Tenstorrent, bringing up generative models and optimizing inference on AI accelerators. I also work with Prof. Meng Tang at UC Merced's GSI Lab on self-consistency and test-time scaling for video reasoning.
Previously, I worked on distributed multi-agent systems with Prof. Baochun Li's iQua Research Group, interpretable mental health prediction with the Kundur Research Group, and LLM-based teaching assistants at the National University of Singapore.
Download my CV (Updated: September 2026)
Education
University of Toronto
3.91/4.0 GPA
Bachelor of Applied Science in Computer Engineering
Work & Research Experience
Tenstorrent
Machine Learning Engineer Intern
Brought up DiffusionGemma, Wan2.2, and MiniMax-H3 on Tenstorrent hardware in tt-metal, from initial porting to functional inference. Optimized compute kernels and integrated vLLM for high-throughput LLM serving.
University of California, Merced
Undergraduate Researcher
Working with Prof. Meng Tang at GSI Lab. First-authored work on self-consistency for diffusion-based video reasoning: training-free test-time scaling raised visual search accuracy from 48.4% to 99.0%, while early readout cut denoising steps by over half. Consensus-based Rejection Fine-Tuning with LoRA improved segmentation gIoU from 0.353 to 0.513 and 4×4 maze success from 72.0% to 84.0%.
University of Toronto
Undergraduate Researcher
Worked with Prof. Baochun Li’s iQua Research Group on reliable and efficient infrastructure for distributed multi-agent systems. Designed and evaluated workflow orchestration for fault tolerance, recovery, and resource efficiency; contributed to a paper under review at IEEE INFOCOM 2027.
University of Toronto
Undergraduate Researcher
Developed a KAN-based multitask model with the Kundur Research Group for depression and anxiety prediction from wearable data, achieving 0.731 macro-AUROC. Built a symbolic surrogate retaining 98.2% accuracy with interpretable rules, with work presented at the NeurIPS 2025 BrainBodyFM Workshop.
National University of Singapore
Research Assistant Intern
Fine-tuned Qwen2 and Llama 3 with LoRA as AI teaching assistants for programming contests. Deployed the models on Alibaba Cloud ECS with vLLM behind a FastAPI service with asynchronous batch processing.
Publications
Honors
- Dean's Honours List (all semesters).
- Top 30 ECE Student, University of Toronto, 2023–2024.