01

About

I am a master's student in the School of Computer Science and Technology at the Harbin Institute of Technology, Shenzhen (expected March 2027), advised by Prof. Wenjie Pei. My work spans two connected mainlines: mobile / GUI agents and multimodal large models — covering data synthesis, pre/post-training (SFT & RL), efficient inference, and generative evaluation.

Since March 2026 I have been with the Tencent Hunyuan team (Qingyun Talent Program), researching mobile-agent algorithms — data-centric trajectory synthesis, agentic RL, and a hybrid-action agent framework. From May 2025 to March 2026 I was at Tencent WeChat AI (WXG), working on large-scale multimodal data engineering and image generation / evaluation.

I have published 5 first-author (incl. co-first) papers across CVPR ×2, ICLR ×1, ECCV ×1, and arXiv. If my research resonates with you, I'd love to connect.

DegreeMSc, Electronic & Information Eng.
AffiliationHIT, Shenzhen
UndergradCS · GPA 3.94/4.0 · Top 5%
HonorsNational Scholarship 2024–25
NowTencent Hunyuan
Based inShenzhen, China
02

Experience

Tencent · TEG Hunyuan QINGYUN TALENT
2026.03 – Present
Mobile / GUI Agent — Algorithm R&D
  • Data-centric synthesis pipeline (lead). Built a multi-function trajectory pipeline (functionality extraction → instruction synthesis → CoT completion → reflection & failure-rewrite reuse), reverse-synthesizing instructions from ~1.05M functionality points across 113 apps (68 Apps + 45 mini-programs) into a million-scale action-trajectory corpus. A 9-stage cleaning stack lifted trajectory retention 30% → 62.5% and CoT accuracy +6%.
  • SFT & agentic RL. Worked across the full base-VLM (A3B) agent pipeline — mid-train (~50B action + 300B corpus), SFT (8.6B aligned data), and RL — exploring trajectory + QA + grounding source/difficulty mixing; shipped offline GRPO RL with verifiable rule-based rewards and contributed to online RL (rubric reward model + multi-env rollout). AndroidWorld success rate 60.3% → 83.2%.
  • Long-horizon decision & memory. Designed a five-stage structured decision schema (current-state / long- & next-plan / action / expected-result) with history compression, supporting tens-of-step trajectories under a 32K context, plus a dead-loop reflection mechanism for multi-turn stability.
  • Hybrid-action agent framework (co-first). Led PhoneHarness, unifying GUI / CLI / MCP actions with deterministic routing and auditable execution traces — 75.0% pass rate, +12.9pt over the strongest non-Harness setting.
Tencent · WXG (WeChat AI)
2025.05 – 2026.03
Multimodal Data Engineering · Vision Encoders & Generative Evaluation
  • Deployed multi-instance Qwen2.5-VL-72B / Qwen3-8B on a 240× P800 GPU cluster; single-handedly built and operated an automated cleaning / quality-filtering / captioning pipeline over 200M image–text pairs feeding multimodal training.
  • Led construction of a 300K-scale image-evaluation dataset and from-scratch T2I base-model training (incl. visual tokenizer / VAE); further built the 1.3M+ CFD benchmark and the CFM evaluation model — owning the full topic-to-acceptance loop, published at CVPR 2026 (Highlight).
03

Research Focus

Mainline A

Autonomous GUI / Mobile Agents

End-to-end agent systems for real devices — large-scale trajectory synthesis, agentic SFT & RL, long-horizon decision making, and hybrid GUI/CLI/MCP action frameworks.

AndroidWorld GRPO / Online RL Trajectory Synthesis Long-horizon Memory Hybrid Action
Mainline B

Multimodal Large Models

Vision encoders and cross-modal contrastive representation, efficient VLM inference via visual-token compression, generative evaluation systems, and scene-text understanding & editing.

Vision-Language Models Token Compression Generative Eval OCR / Scene Text Image Editing
04

Publications

github.com/ZhengyaoFang/WeCromCL
ECCV 2024
WeCromCL: Weakly Supervised Cross-Modality Contrastive Learning for Transcription-only Text Spotting
arXiv · code →
↗ Open repository
ECCV 2024 · co-first author

WeCromCL: Weakly Supervised Cross-Modality Contrastive Learning for Transcription-only Supervised Text Spotting

Jingjing Wu, Zhengyao Fang*, et al.  *equal contribution

Character-level cross-modal contrastive learning trains an OCR vision encoder by modeling appearance consistency between text and image regions, generating pseudo-labels — outperforming CLIP/oCLIP full-image contrast across four benchmarks.

↗ arXiv ⌗ Code
Preprints under review · Mobile-Agent line at Tencent Hunyuan
phoneharness.github.io
↗ Open project page
arXiv 2026 · co-first author

PhoneHarness: Unifying Hybrid GUI / CLI / MCP Actions for Mobile Agents

…, Zhengyao Fang*, et al.  *equal contribution

A hybrid-action mobile-agent framework with deterministic routing and auditable execution traces, reaching 75.0% pass rate (+12.9pt over the strongest non-Harness baseline).

⌂ Project Page ↗ arXiv
phonebuddyai.github.io
↗ Open project page
arXiv 2026 · contributing author

PhoneBuddy: Training Open Models for Agentic Phone Operation

…, Zhengyao Fang, et al.

A joint real-app + mock-environment SFT+RL recipe for mobile-operation agents, lifting AndroidWorld success to 83.2%. I contributed trajectory synthesis & collection, 9-stage cleaning, and SFT/RL.

⌂ Project Page ↗ arXiv
arxiv.org/abs/2607.14548
arXiv 2026
HyMobileAgent: Data–Environment Co-Scaling for Efficient GUI Agents
arXiv →
↗ Open on arXiv
arXiv 2026 · contributing author

HyMobileAgent: Data–Environment Co-Scaling for Efficient GUI Agents

Hy Vision Team, …, Zhengyao Fang, et al.

An efficient GUI-agent recipe co-scaling data and environments — mock-interface synthesis with rejection sampling, video-to-interaction transformation, a million-scale action pipeline across 2000+ device instances, and the PhoneWorld Mock App Factory (34 apps) with dead-loop-aware planning; training spans mid-train, SFT, and RL.

↗ arXiv
05

Contact

Email zhengyaonineve@outlook.com GitHub @ZhengyaoFang Google Scholar Zhengyao Fang