Yu Hu (胡宇)
PhD Student
IntelliArchLab
Rice University
Houston, TX, USA
yh188@rice.edu
github.com/yu44MTX
Education
- 2025 – Present. Ph.D. in Electrical and Computer Engineering, Rice University.
Advisor: Prof. Tony Geng. Lab: IntelliArchLab.
Research Interests
My research focuses on diffusion models, deep-supervised LLMs, and AI security — with a current emphasis on multi-turn jailbreak attacks against aligned language models. Concretely, I am exploring:
- Diffusion model efficiency — dynamic-resolution sampling and trajectory shaping for accelerated denoising.
- Deep-supervised LLMs (DS-LLM) — reasoning architectures with an internal orchestrator that learns where to think, rather than relying on external prompting scaffolds.
- Multi-turn AI red-teaming — attack frameworks against aligned language models, including approaches that use diffusion-style LLMs as the attacker.
Selected Projects
- DyRes — Dynamic-resolution diffusion sampler. Investigated trajectory policies and the “quality cliff” caused by switching resolution near the end of the sampling schedule. Manuscript in revision.
- DS-LLM — Deep-supervised language model with an internal orchestrator. Phase 1 trains ProsQA and GSM8K-Aug on GPT-2 small as a proof of concept; the long-term goal is a DS-native model with no external scaffolding.
- Multi-turn Jailbreak — Studying multi-turn attack strategies (X-teaming, LARGO, STAR) against aligned chat models; exploring whether diffusion-LLM attackers can produce more effective inpaint-style adversarial dialogues.
Publications
In submission.
Skills
- Languages. Python, C/C++, Shell, LaTeX, Typst.
- ML stack. PyTorch, Hugging Face Transformers, JAX.
- Compute. NERSC Perlmutter (project m4705).
- Tooling. Claude Code, Cursor, Zotero, Overleaf.