|
Jingyang Zhang (张静阳)
I'm a Research Scientist at Virtue AI, working on LLM post-training (SFT + RL) for safety and guardrails. Before that, I was at Sciforium, where I worked on pre-training of native multimodal foundation models from scratch.
I received my Ph.D. from Duke ECE in 2024, advised by Prof. Yiran Chen and Prof. Hai (Helen) Li — my dissertation focused on adversarial machine learning and out-of-distribution detection. I obtained my Bachelor's degree in 2019 from the Department of Electronic Engineering at Tsinghua University.
Email /
CV /
Google Scholar /
GitHub /
LinkedIn /
Twitter
|
|
Professional Experience
Aug. 2025 - Present
Virtue AI, San Francisco, CA
Research Scientist - Post-Train, AI Safety
- Owned multiple end-to-end post-train pipelines (SFT + RL) for various guardrails (prompt, action, code vulnerability), including data curation, reasoning trace construction, training, evaluation, and efficiency improvements.
- Led LLM safety evaluation and red-teaming for real-world enterprise agents, delivering actionable failure analysis across large-scale deployments (Fortune 500-level customers and frontier labs).
- Designed and deployed agent emulation and adversarial testing frameworks, including co-developing an internal framework featuring large-scale real-world sandbox environments to study emergent behavior, misuse risks, and failure modes in tool-using LLM agents.
Jan. 2025 - Aug. 2025
Sciforium, Mountain View, CA
Research Scientist - Foundation Models, Pre-Train
- Led design of a purpose-built native multimodal LLM treating text, image, video, and audio as first-class modalities — unified byte-level representation with no modality-specific encoders, targeting coherent cross-modal understanding and generation.
- Designed model architectures and efficient components (non-complex RoPE, optimized KV cache, FP8 GEMM) enabling scalable training and inference across all modalities.
- Built end-to-end pretraining pipelines with scalable data transformation from raw multimodal inputs (text, image, video, audio) to unified byte representations.
- Conducted distributed pretraining across 64 GPUs on 8 nodes using JAX.
|
Education
Ph.D. Duke University, Durham, NC, U.S.
Electrical and Computer Engineering
Aug. 2019 - Dec. 2024
|
|
B.Eng. Tsinghua University, Beijing, China.
Electronic Engineering
Aug. 2015 - Jul. 2019
|
|
|
Open-source Projects
-
OpenClaw-RL (4k+ stars)
Training agents simply by talking through OpenClaw setup. I enabled LoRA training for resource-constrained settings.
-
OpenOOD (1k+ stars)
The largest and most well-recognized codebase for OOD detection on image data. I am one of the main contributors and led the v1.5 release (40+ SOTA methods integration at ImageNet-level scale).
-
lmms-finetune (370+ stars)
A lightweight codebase for fine-tuning various vision LLMs, including LLaVA-Next, LLaVA-OneVision, Qwen-VL etc.
-
VLM-Visualizer (290+ stars)
Visualizing the attention of vision LLMs.
|
Proactive privacy amnesia for large language models: Safeguarding PII with negligible impact on model utility
Martin Kuo*, Jingyang Zhang*, Jianyi Zhang, Minxue Tang, Louis DiValentin, Aolin Ding, Jingwei Sun, William Chen, Amin Hass, Tianlong Chen, Yiran Chen, Hai Li
ICLR'25
[Paper]
[Code]
[Project Page]
|
Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems
Shaokun Zhang, Ming Yin, Jieyu Zhang, Jiale Liu, Zhiguang Han, Jingyang Zhang, Beibin Li, Chi Wang, Huazheng Wang, Yiran Chen, Qingyun Wu
ICML'25 (Spotlight)
[Paper][Code][Project Page]
|
OpenOOD v1.5: Enhanced Benchmark for Out-of-Distribution Detection
Jingyang Zhang, Jingkang Yang, Pengyun Wang, Haoqi Wang, Yueqian Lin, Haoran Zhang, Yiyou Sun, Xuefeng Du, Kaiyang Zhou, Wayne Zhang, et al.
DMLR (Journal of Data-Centric Machine Learning), NeurIPS'23 Workshop on Distribution Shifts (Oral)
[Paper]
[Code]
|
|
Last update: April 2026
|