Wuhan University · Software Engineering · 2027

Start from the problem itself

I am Bohan Su, an undergraduate in Software Engineering at Wuhan University. My current work spans visual representation under appearance variation, small-model structured scientific reasoning, and multi-agent systems for real tasks. I care about methods that become reproducible, analyzable, and deployable systems rather than isolated benchmark gains.

Computer Vision Structured Reasoning Multi-Agent Systems Evaluation Design

Research State

Three active research tracks

Appearance-variation vision, schema-first scientific reasoning, and multi-agent systems are moving in parallel as one coherent portfolio.

Current Output

Papers, systems, and competitions together

An ECCV 2026 submission, PEARL system progress toward EMNLP, a TPAMI survey draft, and national-level systems work define the current output.

Collaboration Fit

Best for reproducible, deployable work

I contribute most on research prototypes and systems projects that need explicit evaluation logic, engineering closure, and clear writing.

Recent Signals

Recent Updates

Research, systems work, and competition projects are moving in the same loop.

2026.04

TPAMI survey main body drafted

§2–§8 of the appearance-variation ReID survey and 231 references are drafted; §1 Introduction and literature alignment are still under revision.

2026.03

ECCV 2026 submission completed

Submitted Reliability-Aware 3D Geometric Injection for Universal Person Re-identification.

2026.03

PEARL pipeline ready; Gate B passed

The graph_spec IR, validator, compiler, issue-list repair loop, and joint-score evaluation pipeline are in place; Gate B local checks passed and Phase 3–5 main experiments are queued.

2025.11

A2A multi-agent system delivered

Built a Holos + A2A workflow with architectural agents, task agents, protocol adaptation, streaming, fallback, and observability for end-to-end demos.

Current Focus

What matters most

  • Turning research claims into explicit evaluation protocols instead of loose benchmark narratives.
  • Separating generation from verification so small models can produce verifiable structure reliably.
  • Building agent systems that can clarify requirements, route work, and expose their own failure modes.

Main Tracks

Research Directions and Papers

I am most interested in topics that connect method design, system behavior, and evaluation logic.

Direction 01

Visual representation under appearance variation

Studying how visual evidence breaks and reorganizes under occlusion, clothing change, cross-modality, and in-the-wild viewpoints in Universal ReID, including UniGeo's reliability-aware 3D geometric injection and a broader survey-level taxonomy.

Direction 02

Schema-first structured scientific reasoning with small models

Decoupling structure learning from syntax generation and diagnosis from repair. PEARL turns ARCHE's scientific reasoning graph extraction into a validator-centered, verifiable training framework for smaller models.

Direction 03

Multi-agent systems for real tasks

Guided mobile interaction, A2A-protocol multi-agent workflows, and modular robot communication, with evaluation protocol, observability, and deployment logic treated as first-class design goals.

Target: EMNLP 2026.03 - present

PEARL: Decoupling Generation from Verification for Small-Model Scientific Reasoning Graph Extraction

For ARCHE's Peircean scientific reasoning graph extraction task, PEARL decouples structure learning from DOT syntax generation and diagnosis from repair. I designed the graph_spec intermediate representation, validator, compiler, issue-list repair loop, and joint-score evaluation pipeline; a 4B Qwen student handles generation, while structural checking and repair are managed externally and trained with five-teacher quality-weighted SFT plus Smart Prompt.

graph_spec IR validator + compiler issue-list repair 5-teacher weighted SFT joint-score target
TPAMI Survey Preparation 2026.03 - present

Appearance Variations in Person Re-Identification: A Survey

Reconstructs the field through a three-branch framework of short-term, long-term, and compound appearance variations, then reorganizes existing methods through a robustness-vs-generalization lens. The long-term branch is further split into clothing state, beyond-clothing, and deployment evolution. §2–§8 and 231 references are drafted, with current attention on evaluation blind spots and metric transferability.

short / long / compound robustness vs. generalization 231 references evaluation blind spots

Hands-on Work

Project Experience

I prefer building systems that are runnable, observable, and iteratively improvable.

Background

Education and Recognition

I keep research, competitions, and systems implementation moving in parallel.

Education

Wuhan University

B.Eng. in Software Engineering, 2023.09 - 2027.06 (expected)

  • GPA 3.74 / 4.0
  • Research interests: visual representation under appearance variations, structured scientific reasoning with small models, multi-agent systems
  • Core courses: Computer Vision, Frontier Technologies in AI, Computer Networks, Operating Systems, Software Engineering, SQA and Testing, Database Systems

Competitions

Selected Competition Achievements

  • National University IoT Design Competition (Huawei Cup): National First Prize, Top 6
  • China Youth Sci-Tech Innovation Challenge: National Second Prize
  • National University Computer System Capability Competition: National Third Prize
  • MCM/ICM: Honorable Mention

Scholarships

Scholarships and University Honors

  • Zheng Geru First-Class Scholarship, ranked 1st in grade defense and sole recipient
  • Wuhan University First-Class Scholarship, twice
  • Merit Student, twice
  • Outstanding Student Cadre, Outstanding Youth League Branch Leader, Future College Outstanding Trainee

Skills and Service

Core Skills and Student Activities

Research work is supported by implementation depth, academic writing, and team coordination.

Core Skills

Engineering and Research Toolkit

  • Python, C/C++, Java; Git and MATLAB
  • Person Re-ID, SMPL body modeling, 3D Gaussian Splatting, experimental design, ablation analysis, academic writing
  • Multi-agent design, orchestration, validation, observability, and application development with GPT, Claude, and Gemini APIs
  • Embedded driver development and protocol design on RISC-V / ESP32 dual-architecture platforms

Student Activities

Representative Service Work

  • Vice Minister of Publicity, Student Party Branch Secretaries Joint Council, Wuhan University (2024.09 - 2025.09)
  • Supported undergraduate Party-building publicity work and contributed to more than 100 published articles
  • Project Team Member, Second Classroom Center, University Youth League Committee (2023.10 - 2024.09)
  • Participated in platform quality monitoring, issue investigation, and follow-through for platform fixes

Working Style

Where I tend to contribute best

Problem framing Benchmark design System prototyping Training pipelines Evaluation loops Research writing

Contact

Open to research collaboration, RA opportunities, and systems-oriented projects

If you are working on visual understanding, structured reasoning, or multi-agent systems, feel free to reach out.