Curriculum vitae · July 2026

Fengji Zhang

PhD candidate · Post-training for general-purpose LLM agents

fengji.zhang [at] my.cityu.edu.hk github.com/zfj1998 Google Scholar · 1.7k+ citations Hong Kong · Fall 2027 · Available Oct 2027

Research profile

PhD candidate in Computer Science at City University of Hong Kong, expecting to graduate in Fall 2027. I work on post-training for general-purpose LLM agents across data, evaluation, reinforcement learning, model behavior, and delivery.

At Alibaba Qwen, I contribute to post-training for Deep Research, adaptive search, and skill use. Earlier work spans reliable search agents, multimodal evaluation, repository context, and execution-guided code generation.

Research experience

Alibaba QwenFoundation Model Team · Research Intern

Mar 2025 — present

Contribute across the post-training loop for general-purpose agents: data construction and analysis, evaluation, SFT/RL teacher models, data and model consolidation, and model delivery.
Work on production-facing Deep Research, adaptive search, and skill-calling behavior, including asynchronous agent workflows.
Parallel academic explorations led to A²Search and AWA-RL.

01.AIPretrain & Multimodal Group · Research Intern

Jun — Dec 2024

Worked on Yi-VL post-training and evaluation for instruction following and visual understanding.
Developed and evaluated Yi-Coder for code generation, completion, and coding-agent settings; studied test-time scaling with long reasoning and execution feedback.
Developed HumanEval-V during this internship as a separate academic project.

Microsoft Research AsiaData, Knowledge & Intelligence · Research Intern

Jun 2022 — May 2023

Proposed CodeT, using generated tests and execution agreement to improve inference-time selection for code generation.
Developed RepoCoder, an iterative retrieval-generation approach for repository-level code completion and cross-file context.

TencentIEG

Jun — Sep 2019

Developed AIOps and internal SaaS tooling for online-game operations.

Selected papers

Preprint · 2026

To Answer or to Abstain: Mitigating Search-Agent Hallucinations via Abstention-Aware Reinforcement Learning

Fengji Zhang, Tianyu Fan, Yuxiang Zheng, Xinyao Niu, Chengen Huang, Jacky Keung, Bei Chen.

AWA-RL learns a controllable answer–abstain policy from query-specific prior capability and on-policy observations.

ICLR · 2026

A²Search: Ambiguity-Aware Question Answering with Reinforcement Learning

Fengji Zhang, Xinyao Niu, Chengyang Ying, et al.

Annotation-free ambiguity resolution with trajectory sampling, evidence verification, and an AnsF1 reward for multiple valid answers.

ACM TOSEM · 2026

HumanEval-V: Systematic Evaluation of Visual Reasoning in Large Multimodal Models for Code Generation

Fengji Zhang, Linquan Wu, Huiyu Bai, et al.

A human-annotated benchmark for testing visual understanding through executable coding tasks with indispensable diagram context.

EMNLP · 2023

RepoCoder: Repository-level Code Completion through Iterative Retrieval and Generation

Fengji Zhang, Bei Chen, Yue Zhang, et al.

Iterative retrieval-generation for cross-file context; more than 10% improvement over in-file completion baselines across RepoBench settings.

ICLR · 2023

CodeT: Code Generation with Generated Tests

Bei Chen*, Fengji Zhang*, Anh Nguyen*, et al.

Generated tests and dual execution agreement for solution selection; 65.8% pass@1 on HumanEval.

Additional work spans code intelligence, multimodal understanding, deep research, software engineering, and responsible code generation. View the complete publication record on Google Scholar ↗

Education

City University of Hong Kong
PhD candidate in Computer Science · 2023–2027 expected

Wuhan University
MS in Computer Science · 2020–2023
BS in Computer Science · 2016–2020

Service & recognition

Reviewer
ICSE ’24, ICLR ’25/’26, ICML ’26, NeurIPS ’26, CVPR ’25, ICCV ’25, ACL ’25/’26, EMNLP ’26

Award
CityUHK Outstanding Academic Performance