01
Content provenance
Can generated content carry a signal without changing its distribution?
Unbiased and distribution-preserving watermarking explores traceability while protecting output quality and model utility.
Zhengmian Hu, PhD · AI Researcher
Trustworthy generation.
Faster inference.
I develop watermarking and sampling methods for tracing generated content and accelerating LLM inference, while studying where output distributions can be preserved—and where trade-offs are unavoidable.
01Provenance
02Generation
03Inference
Research focusResponsible
generative AI
generative AI
PhD · 2024University of Maryland
ICLR · ICML · NeurIPSPeer-reviewed research
Open to collaborationTechnical evaluation & technology transfer
01 / Selected work
One research program, three practical questions.
The through-line is exactness: add useful capabilities while preserving the behavior that makes a probabilistic model valuable.
02
Inference systems
Can language models produce more tokens per step and stay exact?
Accelerated Speculative Sampling reframes generation on a tree space to improve inference efficiency while maintaining the target distribution.
EvidenceICML 2024
03
Joint constraints
What changes when provenance and acceleration meet?
A no-go result identifies an inherent trade-off between maximum watermark strength and maximum speculative-sampling efficiency, then develops methods for either operating point.
EvidenceNeurIPS 2024
02 / From paper to practice
Published research, open to technical evaluation.
Watermarking & provenance
Research collaboration and evaluation for content attribution, model-output traceability, and privacy-aware verification.
Efficient inference
Technical exploration for teams working on latency, throughput, and exact speculative generation in LLM serving systems.
Technology transfer
I welcome private diligence on intellectual-property status and scope before any technology-transfer discussion.
Open to discussionResearch collaborationTechnical evaluationTechnology-transfer diligenceResearch & industry roles
03 / EU policy context
Transparency has moved from principle to implementation.
Article 50 of the EU AI Act has applied to covered systems since 2 August 2026, with a transitional period to 2 December 2026 for the marking of outputs by generative systems already on the market. The European Commission describes machine-readable marking and detectability among certain providers’ obligations; scope and exceptions depend on the use case.
European Commission · Article 50 guidance04 / Publications
Selected, peer-reviewed work.
01Unbiased Watermark for Large Language ModelsA distribution-preserving route to tracing and attributing model outputs.ICLR 202402Accelerated Speculative Sampling Based on Tree Monte CarloFaster inference while maintaining the target model’s original distribution.ICML 202403Inevitable Trade-off between Watermark Strength and Speculative Sampling Efficiency for Language ModelsA theoretical account of what can—and cannot—be optimized at once.NeurIPS 202404A Resilient and Accessible Distribution-Preserving Watermark for Large Language ModelsDetectable without model API access and resilient to moderate token changes.ICML 202405Enhancing privacy in biosecurity with watermarked protein designPrivate local verification and robust traceability for designed sequences.Bioinformatics 2025
05 / About
I work across machine-learning theory and generative-AI systems.
I received my PhD in Computer Science from the University of Maryland in 2024, advised by Heng Huang. My doctoral research connects theoretical foundations and algorithm design; recent publications focus on watermarking, speculative sampling, and their trade-offs.
This site brings together published research and the areas where I welcome collaboration: rigorous research, technical evaluation, technology transfer, and research roles.
06 / Contact
Have a hard problem at the boundary of trust and efficiency?
I’m open to research collaborations, technical evaluations, technology-transfer conversations, and selected research or industry opportunities.
[email protected]