Agentic AI
Modular systems that coordinate specialized language agents.
Computer Science · NLP · Agentic AI
Renxiang
Wang
王仁翔
I build language agents that reason, retrieve, and collaborate more efficiently.
YOUR
PHOTO
PHOTO
01 / ABOUT
Researching capable, efficient language agents.
I am Renxiang Wang (王仁翔), a Computer Science undergraduate at Zhejiang Sci-Tech University, graduating in June 2026. My research spans modular multi-agent LLM systems, reasoning-driven retrieval, natural language reasoning, reinforcement learning for language agents, and evaluation under uncertain rewards.
Across my work, I ask a practical question: how can we allocate knowledge, reasoning, and compute to the right stage of an AI workflow?
RESEARCH INTERESTS
Reasoning & Retrieval
Planning information needs before evidence acquisition.
Efficient Inference
Routing models and compute according to workflow roles.
02 / ONGOING PROJECTS
Current work
Ongoing · Medical AI
Clinician-Aligned Multimodal Spine Imaging
Building a clinically grounded system that organizes X-ray, CT, and MRI at the study level, combines specialized perception models with reproducible measurements, and generates traceable dense observations and report drafts.
Current work aligns outputs with clinicians through blinded A/B review, error labels, and revisions, using preference optimization while preserving localization accuracy and safety constraints.
Multimodal VLMs · Evidence grounding · Clinician preference alignment
Ongoing · Agent Systems
Large-Scale Agent Skill Transfer
Studying whether relative routing and communication skills learned on one model ladder can transfer to another without fine-tuning the underlying models.
Experiments span CF and FanOutQA workflows at multiple agent scales, comparing capability and communication transfer with training-free baselines under held-out quality, cost, message-budget, and negative-transfer metrics.
Routing · Communication DAGs · Training-free transfer
03 / SELECTED PUBLICATIONS
Research outputs
01
IJCNLP-AACL 2025 · Oral
Documentation Retrieval Improves Planning Language Generation
02
IEEE VISxGenAI 2025
A2P-Vis: An Analyzer-to-Presenter Agentic Pipeline for Visual Insights Generation and Reporting
03
ACL Findings 2026
Scaling Unverifiable Rewards: A Case Study on Visual Insights
04
Preprint · AAAI 2027 under review
Beyond Tier Labels: Role- and Deployment-Dependent Model Substitution in Multi-Call LLM Workflows
04 / EXPERIENCE
Research journey
May 2026 — Present
Research Assistant · Virginia Tech
Supervised by Prof. Jiaming Cui
Developing EvoCap, a trace-driven routing framework for multi-agent LLM workflows. The work matches all-strong performance with roughly 10% strong-model calls and reduces test-time API cost by 73.9%.
Apr 2025 — May 2026
Research Assistant · Drexel University
Supervised by Prof. Li Zhang
Built a modular PDDL generation pipeline combining documentation retrieval, code generation, and error refinement, improving syntactic correctness from 0% to 90%+ and semantic correctness from 0% to 80%+.
May 2025 — May 2026
Research Assistant · University of Minnesota
Supervised by Prof. Dongyeop Kang
Worked on multi-agent visualization and verification. Proposed selective test-time scaling to improve insight quality while reducing variance by 25%.
Aug 2024 — Jun 2026
Research Assistant · Zhejiang Sci-Tech University
Supervised by Prof. Zhiyi Luo
Studied reasoning-driven retrieval and multi-span question answering, including TROVE, TOAST, CLEAN, and SIGMA-CoT across English and Chinese benchmarks.
LET'S CONNECT