CUA Agent Trajectory Pipelines
Constructing high-quality human demonstrations and automated rollouts across desktop, web, and game environments, including action schemas, validation, filtering, and conversion.
Senior Researcher at Tencent Hunyuan
Building CUA agents that learn from trajectories, tools, and environments.
I am a senior researcher at Tencent Hunyuan. I received my Ph.D. from the School of Computer Science and Engineering, South China University of Technology, advised by Prof. Qianli Ma. My current work focuses on CUA agents, including human and automated trajectory construction, supervised fine-tuning, and reinforcement learning for game-play agents.
Research Focus
Towards general-purpose CUA agents that can get real work done
SFT for Actionable Agents
Constructing productivity data from real work scenarios, designing data mixtures for CUA agent SFT, and applying step-level loss masking so the model learns decision-relevant action traces without overfitting to redundant trajectory information.
Game-play RL
Building Nexus-based game-play sandboxes and reinforcement learning loops where CUA agents interact with environments, receive task progress feedback, and improve through executable rollouts.
Experience
Research homes and collaborators
Jan 2026 - Present
Tencent Hunyuan, Senior Researcher
CUA agent research, including human and automatic trajectory data construction pipelines, supervised fine-tuning, and game-play reinforcement learning.
Apr 2025 - Dec 2025
Tencent Hunyuan X, Qingyun Intern
Reinforcement learning and post-training for large reasoning models, including math reasoning and Hunyuan model reports.
Jul 2024 - Apr 2025
Microsoft AI, Research Intern
Early-stage training and evaluation for WizardLM-2, followed by leading WarriorCoder, a data-flywheel project where a target model learns from expert battles to augment code generation.
Jul 2023 - Jul 2024
Alibaba Tongyi Lab, Research Intern
Factuality, preference optimization, and correction methods for LLMs, including contrastive preference optimization and backtracking correction.
Apr 2023 - Jul 2023
Tencent AI Platform Department, Algorithm Intern
Worked on an intelligent COVID-19 epidemiological investigation data collection project, covering algorithms such as NER, object detection, OCR, and address recommendation, substantially improving the efficiency of contact-tracing information collection.
Selected Publications
Recent work across agents, post-training, and NLP
AgentMath: Empowering Mathematical Reasoning for Large Language Models via Tool-Augmented Agent
H. Luo, H. Feng, Q. Sun, C. Xu, K. Zheng, Y. Wang, T. Yang, H. Hu, Y. Tang
AgentMath integrates language-model reasoning with code interpreters, converting natural-language chain-of-thought into structured tool-augmented trajectories for SFT. It also trains an agentic RL loop that interleaves generation and code execution, helping models learn stronger tool-use and self-correction strategies for math.
Hunyuan-TurboS: Advancing Large Language Models through Mamba-Transformer Synergy and Adaptive Chain-of-Thought
Tencent Hunyuan Team
Hunyuan-TurboS is a large hybrid Transformer-Mamba MoE model built for efficient long-context reasoning and deployment. The report combines adaptive long/short chain-of-thought, multi-round deliberation, and large-scale RL to balance strong capability with lower inference cost.
WarriorCoder: Learning from Expert Battles to Augment Code Large Language Models
H. Feng, P. Zhao, Q. Sun, C. Xu, F. Yang, L. Wang, Q. Ma, Q. Lin, S. Rajmohan, D. Zhang, Q. Zhang
WarriorCoder builds a data flywheel for code LLMs from expert-model battles rather than static datasets or proprietary-only augmentation. An arena lets strong code models challenge one another and be judged impartially, producing diverse training data for stronger code generation.
Training Large Language Models for Retrieval-Augmented Question Answering through Backtracking Correction
H. Feng, J. Zheng, Q. Ma
Backtracking Correction trains RAG QA models to recognize when a generation drifts from retrieved evidence and revise the answer instead of continuing an unfaithful path. The method turns correction behavior into trainable trajectories for more faithful retrieval-grounded answering.
Improving Factual Consistency of News Summarization by Contrastive Preference Optimization
H. Feng, Y. Fan, X. Liu, T.-E. Lin, Z. Yao, Y. Wu, F. Huang, Y. Li, Q. Ma
CPO targets the more subtle hallucinations that modern LLMs make in news summarization, such as invented causal links or overgeneralized details. It separates faithful and fake generation tendencies with contrastive preference training, improving factual consistency and hallucination awareness.
Well Begun Is Half Done: An Implicitly Augmented Generative Framework with Distribution Modification for Hierarchical Text Classification
H. Feng, J. Yan, J. Liu, J. Zheng, Q. Ma
IMAGE addresses hierarchical text classification when label distributions are highly imbalanced and generative decoding is vulnerable to early errors. It implicitly augments sample distributions and transfers label-space knowledge to tail classes, making predictions less biased and more robust.
Preserving Commonsense Knowledge from Pre-trained Language Models via Causal Inference
J. Zheng, Q. Ma, S. Qiu, Y. Wu, P. Ma, J. Liu, H. Feng, X. Shang, H. Chen
This work revisits fine-tuning PLMs for commonsense QA from a causal inference perspective, showing that vanilla fine-tuning can overfit target data and lose useful causal effects from pre-training. It introduces a unified objective that learns task-specific knowledge while preserving transferable commonsense knowledge from PLMs.
Perturbation-Based Self-Supervised Attention for Attention Bias in Text Classification
H. Feng, Z. Lin, Q. Ma
This paper studies attention bias in text classification, where learned attention weights can become weak explanations or focus on misleading tokens. It uses perturbation-based self-supervision to guide attention toward decision-relevant evidence, improving robustness and interpretability.
Joint Constrained Learning with Boundary-Adjusting for Emotion-Cause Pair Extraction
H. Feng, J. Liu, J. Zheng, H. Chen, X. Shang, Q. Ma
JCB addresses emotion-cause pair extraction under severe imbalance among emotion clauses, cause clauses, and paired relations. It injects prior constraints into learning and adjusts classifier decision boundaries so subtasks cooperate instead of being optimized independently.
It's Better to Teach Fishing than Giving a Fish: An Auto-Augmented Structure-aware Generative Model for Metaphor Detection
H. Feng, Q. Ma
AAAS reframes metaphor detection as a generative keyword-extraction task with sentence-structure awareness. It adds an auto-augmentation strategy that expands limited data without external resources, improving contextual modeling for metaphorical word detection.