CV

AI Research Intern

TikTok Search (ByteDance)

Agentic Search and Tako: scenario-aware agent policies, evidence-driven system iteration, and ongoing model-training design.

• Led Auto-Skill policy iteration: failure diagnosis, bounded strategy changes, versioned experiments, and evaluation of the delivered search experience.
• Developed scenario-specific prompts while preserving shared planning and response contracts.
• Ongoing data work - faithful trajectories that distinguish model-visible evidence, agent actions, runtime repairs, and final rendered outcomes.
• Ongoing training and evaluation design - compare SFT, reinforcement learning, and on-policy distillation; evaluate agent behavior and product quality separately.

AI Research Intern

TikTok Shop (ByteDance)

Led FIRE, EvoA, and APA from design to production within five months, working with business and infrastructure partners on integration and rollout.

• FIRE - designed a reusable skill-based agent harness with semantic routing, progressive context loading, and governed tool execution for e-commerce workflows.
• EvoA - built an autonomous model-iteration loop combining diagnosis, pilot experiments, validation gates, and reusable experiment memory; reduced the human audit rate by approximately 4.6% relative in a controlled online experiment.
• APA - built evidence-driven risk discovery using joint content / product signals, scoped expansion, deduplication, and provenance-aware review; expanded candidate throughput approximately 17-fold in workflow measurements.
• Owned architecture, implementation, experimentation, evaluation, and cross-team integration.

Ph.D. Researcher

Show Lab, National University of Singapore

Research on grounded agent recovery, reliable supervision, evolving state, and continual model adaptation.

• ReGRPO - grounded reflection and corrective tool use; accepted at ECCV 2026.
• SGRC, StateTrackBench, and PathComp - first-author submissions under review at ICLR 2027.
• POU-OPD - first-author submission under review at AAAI.
• Four first-author or co-first-author conference papers across ICLR, IJCAI, AAAI, and ECCV; TaCA is an additional preprint.

AI Research Intern

Tencent ARC Lab - Compatible Representation Learning

Backward-compatible representation learning and model upgrades for large-scale image and video retrieval systems.

• First-author / co-first-author work at ICLR 2022, IJCAI 2022 (long oral), and AAAI 2023, plus the TaCA preprint.
• Designed compatible model upgrades that accelerate production hot updates and reduce feature-refresh cost for image and video retrieval.
• Built distributed contrastive / CLIP-style training pipelines and regression-alleviating compatible learning methods.
• Recipient of the Tencent Technology Breakthrough Award and SZCCF Science and Technology Award.

AI Research Intern

Tencent ARC Lab - Cross-Modality Video Understanding

Multimodal models for video-text retrieval and temporal grounding on large-scale video-language datasets; built the PyTorch data / training / evaluation codebase for the group’s multi-modal research.

Ph.D.

National University of Singapore (NUS)

Computer Science

• Advisor - Prof. Mike Zheng Shou (Show Lab).
• Research on LLM agents, multimodal understanding, and continual / compatible learning.
• Expected completion - January 2027.

M.Eng.

Tsinghua University (THU)

Computer Science and Technology

• Advisor - Prof. Chun Yuan.
• Focus on compatible representation learning and cross-modal video understanding.
• Excellent Master’s Graduate in Beijing; Outstanding Master’s Thesis.

B.Eng.

East China University of Science and Technology (ECUST)

Information Engineering

• Cumulative rank 1 / 92.

Tencent Technology Breakthrough Award

Tencent

For Hot-Refresh Model Upgrades.

SZCCF Science and Technology Award

Shenzhen CCF

For efficient compatible model upgrades.

Excellent Master Degree Graduate in Beijing & Outstanding Master's Thesis

Beijing Municipal Education Commission / Tsinghua University

Annual College Personage Award

ECUST

Highest student honor at ECUST.

National Scholarship for Undergraduates (twice)

Ministry of Education of China

2016 and 2017.

ReGRPO: Reflection-Augmented Policy Optimization for Tool-Using Agents

ECCV 2026 (accepted)

First author. Grounded reflection and corrective tool use.

TaCA: Upgrading Your Visual Foundation Model with Task-agnostic Compatible Adapter

arXiv preprint, 2023

First author. Preserve downstream interfaces during visual backbone upgrades.

Darwinian Model Upgrades: Model Evolving with Selective Compatibility

AAAI 2023

Co-first author. Selective backward compatibility and forward adaptation.

Towards Universal Backward-Compatible Representation Learning

IJCAI 2022 (long oral)

First author. Compatible upgrades across different classes and training settings.

Hot-Refresh Model Upgrades with Regression-Alleviating Compatible Training in Image Retrieval

ICLR 2022

First author. Incremental gallery refresh with compatible model upgrades.

Selective Partition Repair for Group-Relative Policy Optimization

Under review at ICLR 2027

First-author submission (SGRC).

StateTrackBench: Executable Trajectories for Locating Where Language Models Lose Track of Evolving State

Under review at ICLR 2027

First-author submission.

Path Compatibility: Persistent Anchors for Continual VLA Fine-Tuning

Under review at ICLR 2027

First-author submission (PathComp).

Keep or Reset? Measuring Update Utility for On-Policy Distillation of Coding Agents

Under review at AAAI

First-author submission (POU-OPD).