CV - Yuxin Wang

Education

M.S.E. in Data Science, Johns Hopkins University, Jan 2026 – Present
Department of Applied Mathematics and Statistics, Baltimore, MD · GPA: 3.58 / 4.0

B.E. in Data Science and Big Data Technology, Tongji University, Sep 2019 – Jun 2023
College of Electronics and Information Engineering, Shanghai, China · GPA: 3.55 / 4.0
Key Courses: C/C++/Python Programming, Operating System, Data Mining, Data Structure and Algorithm Design, Big Data Management Systems, Cloud Computing, Machine Learning, Artificial Intelligence

Research Experience

Self-Evolving Medical Agent via Counterfactual Reasoning, Mar 2026 – Present
Advisor: Dr. Xing Han, Johns Hopkins University

Designed and implemented an end-to-end self-evolving pipeline for ICU clinical prediction.
Curated a 300-example counterfactual dataset from MIMIC-IV; fine-tuned Qwen2.5-7B to improve counterfactual quality.
Confidence-gated memory injection improved AUROC from 0.736 → 0.804, accuracy 0.727 → 0.840, and F1 0.349 → 0.478 on in-hospital mortality prediction.
Extending the framework to multitask ICU prediction (mortality, length-of-stay, phenotyping).

Loss Landscape Analysis of Scientific Machine Learning Models, Mar 2025 – May 2026
Advisor: Prof. Yaoqing Yang and Pu Ren, Dartmouth College

Developed measurement of mode connectivity in PINNs; characterized optimization regimes based on Hessian, CKA, and mode connectivity.
Conducted systematic comparisons of soft and hard constraint methods (Fourier parameterization, penalty method, Augmented Lagrangian Method, Trust-Region SQP) to analyze performance across different regimes.
Identified and explained several pathological phenomena in SciML that differ from those observed in computer vision, from a loss landscape perspective.
Paper accepted at ICML 2026.

Low-light Image Enhancement Based on Attention Mechanism, Mar 2023 – Jun 2023

Reproduced Illumination Adaptive Transformer and Multi-Axis MLP; conducted experiments on LOL and SICE datasets.
Optimized deep convolution enhancement module in the local branch by incorporating SWIN attention.
Attempted to incorporate black level correction transformation matrices from ISP into the global branch.
Project received 50+ forks on Kaggle; achieved PSNR of 0.65 and SSIM of 20.45.

Professional Experience

Subject Matter Expert – AI, Johns Hopkins University, Jun 2026 – Present, Baltimore, MD
Advisor: Prof. Benjamin Van Durme and Dr. William Walden

Developed a multi-agent deep-research framework for automated scientific claim verification.
Improved memory system with compressed context rendering to mitigate token/context blow-up during long-horizon web research.
Implemented a dual-track LLM-as-judge pipeline evaluating final reports and execution traces via claim-specific obligation maps and G-Eval-style staged rubrics.

Software Development Engineer, Beijing Dajia Internet Information Technology (Kuaishou), Jul 2023 – Mar 2025, Beijing, China

Built performance analysis, monitoring, and traffic replay tools for large-scale mobile systems, enabling rapid diagnosis of memory and performance issues.
Maintained and developed UI components (images, animations, listviews) to support large-scale events on Kuaishou, including Singles’ Day and Chinese New Year.
Designed ETL pipelines and dashboards to analyze JSVM memory usage, reducing issue attribution time from hours to minutes.
Identified and resolved performance and memory inefficiencies in layout, image, animation components and property export; reduced peak memory usage by 56% and rendering time by 11% in dynamic stages for e-commerce and local services scenarios.

Publications

Counterfactual Self-Evolving Agents for Evidence-Grounded Reasoning

Xing Han, Yuxin Wang, Chen Chen, Wei Dai, Gautham Krishna Gudur, Shijun Li, Hsing-Huan Chung, Gregory D. Hager, Joydeep Ghosh, Paul Pu Liang, Suchi Saria. (2026). "Counterfactual Self-Evolving Agents for Evidence-Grounded Reasoning." arXiv preprint arXiv:2609.32870.

Unveiling Multi-regime Patterns in SciML: Distinct Failure Modes and Regime-specific Optimization

Yuxin Wang, Yuanzhe Hu, Xiaokun Zhong, Xiaopeng Wang, Haiquan Lu, Tianyu Pang, Michael W. Mahoney, Yujun Yan, Pu Ren, Yaoqing Yang. (2026). "Unveiling Multi-regime Patterns in SciML: Distinct Failure Modes and Regime-specific Optimization." ICML 2026.

Yuxin Wang

CV