CV

Education

• M.S. in Artificial Intelligence, Fudan University, 2025.09 – 2028.06 (expected)
  • School of Data Science, Knowledge Works Lab
  • Admitted via recommendation (推免)
  • Core courses: NLP & LLM, Multimodal LLMs, Frontier AI Algorithms, Optimization Theory, Deep Learning
• B.Eng. in Computer Science (Honors Class), Hangzhou Dianzi University (Zhuoyue Honor College), 2021.09 – 2025.06
  • GPA: 4.59 / 5.0 (Top 0.6% of major)
  • Outstanding Graduate of HDU

Honors & Awards

• National Scholarship (国家奖学金)
• Provincial Government Scholarship (省政府奖学金)
• Outstanding Graduate of HDU (校优秀毕业生)
• Lead PI, National-level College Student Innovation Project (国家级大创负责人)
• CUMCM (China) — National Second Prize & Provincial First Prize
• MCM/ICM — Meritorious Winner (M Prize)
• Huawei Software Elite Challenge — Third Prize
• 10+ additional competition awards
• Member of HDU ACM Training Team

Research Experience

Locality-aware Diffusion Language Modeling  ·  Sole first author  ·  2025.10 – 2026.03

• Systematically studied the trainability of Masked Diffusion LMs vs. autoregressive LLMs across structured generation tasks, exposing how inductive bias interacts with task dependency structure.
• Designed three controlled tasks — in-context linear regression, star-graph path-finding, and Sudoku — covering local exact binding, reverse planning, and global constraint satisfaction.
• Proposed two locality-aware blockwise diffusion architectures: Scatter (“intra-block AR + inter-block synchronous update”) and Jigsaw (“intra-block AR + inter-block entropy-guided dynamic programming”), unifying local ordering with global iterative refinement.
• Findings: AR is better suited for local binding & sequential generation; Diffusion is better suited for global planning & constraint satisfaction. Provides empirical foundations for dLLM architecture design and future Agent planning research.

STM: Spatio-Temporal Distance and Frame-Based Dynamic Graph Fraud Detection  ·  First author (CCF-B), Invention Patent  ·  2023.04 – 2024.08

• Identified that conventional GNN-based fraud detectors fail to fully exploit spatio-temporal information on dynamic graphs.
• Proposed STM, combining graph condensation, distance-aware intra-frame aggregation, and difference-aware inter-frame aggregation to capture both short- and long-range temporal–spatial signals.
• Achieved SOTA on multiple dynamic-graph fraud-detection benchmarks. Outcome: CCF-B publication and granted invention patent.

Semantic Diffusion Language Modeling (SemDLM)  ·  Second author

• Studied how the noising kernel design in discrete diffusion LMs affects training stability and generation quality; proposed a unified bias–variance trade-off framework.
• Proposed SemDLM: semantic-neighborhood diffusion + shared refresh branch + token-frequency prior correction reduces training bias/variance and strengthens sampling-time error correction.
• Result: 27.19 Test PPL on LM1B, outperforming multiple discrete diffusion baselines.

Industry Experience

Research Intern — AI4S / Multimodal LLM, Alibaba  ·  2025.10 – Present

• Direction: protein/small-molecule unified multimodal foundation models, protein generation/understanding evaluation, and Protein Agent post-training.

• Project I — ProtText-Eval: A Unified Evaluation Framework for Text-Guided Protein Design
  • Solely responsible for designing and shipping the full evaluation infrastructure.
  • Built a 4-tier, 26-metric evaluation suite covering biological-plausibility filtering, instruction alignment, and task-specific capabilities.
  • Designed the modular intermediate protocol ProteinSample → Adapter → ProteinPrediction, decoupling models / metrics / datasets for plug-and-play extension.
  • Engineered PDBPool with MD5 dedup + persistent caching, dramatically improving structure-prediction throughput.
  • Integrated 5+ baseline models and ran 8+ benchmark evaluations; supports both generation and editing evaluation paradigms.
• Project II — Protein-Function Understanding Agent (ReAct + RL)
  • Core designer/developer of agent architecture, tool system, and the RL training pipeline.
  • Built a multi-turn Thought → Action → Observation rollout engine with a Teacher Backend abstraction supporting both local heuristics and external LLM teacher models.
  • Built a registry/execution system covering 18 bioinformatics tools; ported 6+ core tools (mmseqs2, hmmscan, tmbed, signalp, …) to a native backend; designed a layered dependency mechanism to remove implicit tool coupling.
  • Built end-to-end data closure on Alibaba’s ROLL & ROCK frameworks: rollout → trajectory export → SFT/RL config generation → dual evaluation (ROCK tool-call metrics + ProtText biological plausibility).
  • Designed a multi-dimensional reward (answer quality, tool selection, reasoning quality, execution efficiency) for GRPO RL training.
• Project III — Unified Multimodal Protein Foundation Model (Understanding + Generation)
  • Co-designed protein-modality training strategy and the SFT / inference pipeline.
  • Built around a MoT (Mixture-of-Transformer) backbone supporting bidirectional protein ↔ text modeling.
  • Three-stage training: Stage 1 — protein pretraining on UniRef50 (~59M sequences); Stage 2 — protein–text alignment on SwissProt; Stage 3 — SFT for QA / caption / design.
  • Implemented protein/text mixed packed sequences, unified next-token loss (text CE + protein CE), and token-level MoT routing.
  • Built protein understanding & reasoning capabilities: protein + question → answer, sequence-likelihood scoring, and ProteinGym evaluation.
  • Implemented decoupled two-stage training (Stage 1 → Stage 2 checkpoint init) with large-scale distributed training support.
• Stack: Python, Linux, PyTorch, Transformers; Accelerate, DeepSpeed; FlashAttention-2, vLLM; Docker, SLURM, WandB; verl / ROLL, GRPO.

Algorithm Engineer Intern — Intelligent Driving (VLA Dept.), Tianyi Transportation Tech  ·  2025.04 – 2025.07

• Co-designed and shipped an automated 4D pre-labeling pipeline for autonomous-driving data, integrating multiple open-source models for map/road prediction, 3D detection, semantic generation, and Occupancy prediction.
• Owned debugging, adaptation, and integration across heterogeneous open-source algorithms; resolved I/O-format / inference-flow / result-alignment incompatibilities to deliver a unified labeling chain.
• Improved pre-labeling efficiency and multi-task coordination via toolchain & automation work, providing high-quality data for large-scale AD training.
• Stack: PyTorch, CUDA, Linux, Docker, OpenPCDet, MMDetection3D, OpenCV.

Skills

• Languages: C++, Python, Matlab
• ML / DL: PyTorch, Transformers, Accelerate, DeepSpeed, FlashAttention-2, vLLM
• RLHF: verl, ROLL, GRPO
• Engineering: Docker, SLURM, WandB, Git, Linux, GPU cluster operations
• AI Coding Tools: Claude Code, OpenClaw — awarded department’s Top Code Contributor of the Month and presented internal AI-Coding talks
• Math foundation: Mathematical Analysis, Linear Algebra, Operations Research — all 90+ at HDU; HDU ACM training-team member

Publications

Locality-aware Diffusion Language Modeling

Yuxiang Wang, et al. "Locality-aware Diffusion Language Modeling." arXiv:2604.24832, 2026.

Semantic Diffusion Language Modeling

STM: Spatio-Temporal Distance and Frame-Based Dynamic Graph Fraud Detection

Yuxiang Wang, et al. "Spatio-Temporal Distance and Frame-Based Dynamic Graph Fraud Detection." Springer, 2024. https://doi.org/10.1007/978-981-95-3906-2_3