About
Haoran (Harry) Dai is a PhD candidate in Computer Science at the Illinois Institute of Technology, advised by Professor Binghui Wang, and a Research Scientist at Quiver AI. His research examines the security and efficiency of modern AI systems. On the security side, he develops and evaluates backdoor attacks against text-to-image and multimodal diffusion models and vision-language models, together with defenses against them. On the efficiency side, he studies reasoning models and large language models, focusing on test-time inference scaling and the attention and quantization behavior underlying their inference stability. At Quiver AI, he works on SVG generation systems, spanning inference optimization, model post-training, data curation, and the internal tooling that supports them. He is based in Chicago, Illinois.
Research Interests
- Reasoning Models and Test-Time Scaling
- Efficient LLM Inference
- Vector Graphics (SVG) Generation
- Diffusion Model Safety
- Vision-Language Model Safety
News
- May 2026Attention Sinks and Outliers in Attention Residuals released on arXiv.
- Apr 2026TIDES presented as a poster at the ICLR 2026 Workshop on Efficient Spatial Reasoning.
- Mar 2026When One Modality Rules Them All accepted at the ICLR 2026 Workshop on Principled Design for Trustworthy AI.
- Feb 2026VectorGym and WildSVG, two benchmarks for SVG generation, released on arXiv.
- Oct 2025Joined Quiver AI as a Research Scientist, working on SVG generation systems.
- Aug 2025Practical, Generalizable and Robust Backdoor Attacks on Text-to-Image Diffusion Models released on arXiv.
- Aug 2024Started the PhD in Computer Science at the Illinois Institute of Technology, advised by Professor Binghui Wang.
Newest Publications
