AutoResearch & Self-Evolving Agents

Bingchen Huang 黄丙晨

Agent Algorithm Engineer · Meituan

I focus on AutoResearch and self-evolving agents: building agents that carry out research and improve through experience. I initiated and lead Dr. Pie, an enterprise research agent combining research skills, execution harnesses, and skill evolution.

My earlier work focused on large-model post-training, multimodal in-context learning, and continual learning. I received my M.Sc. from the FVL Lab, Fudan University, advised by Prof. Yu-Gang Jiang and Prof. Zhineng Chen, and my B.Eng. from East China University of Science and Technology.

Google Scholar GitHub LinkedIn CV · 中文 PDF
Portrait of Bingchen Huang

§ 01 News

§ 02 Research Interests

AutoResearch

Composable research skills and agents that connect problem formulation, experimentation, evaluation, and scientific writing.

Self-Evolving Agents

Learning from execution traces and failures to improve skills through reflection, offline evolution, and task-level validation.

Long-Horizon Reliability

Execution harnesses, persistent state, process verification, and recovery for research workflows that run over days.

Current project · Initiator & Lead

Dr. Pie Enterprise AutoResearch Agent

2026 — Present

I initiated Dr. Pie and built its early research skills, core harness, self-evolution mechanisms, and desktop application. I now lead its technical direction, development, and benchmark evaluation.

The system combines composable research workflows with state management and recovery for long-running tasks. It learns from execution experience through offline skill evolution, with validation before adopting changes.

Used internally at Meituan by over a thousand users, including in company-level model research workflows.

Earlier work: large-model post-training (SFT / RL), multimodal in-context learning, domain embeddings, and continual learning.

§ 03 Publications

* Equal contribution (co-first authors). Published papers and preprints; ongoing work is listed separately below. See Google Scholar for the full list.

  1. arXiv 2026

    Retrieved Images as Visual Thought: Training-Free Multimodal In-Context Learning for the Open-vs-Closed Gap

    Bingchen Huang, Zhiling Wang, Yifu Chen, Yuanchao Du

    ReVisIT · Retrieved image exemplars as visual thought for multimodal in-context learning.

    arXiv PDF

  2. arXiv 2026

    DRDN: Decoupled Representation Dynamic Network for From-Scratch ViT Class-Incremental Learning

    Bingchen Huang, Yifu Chen, Zhiling Wang, Yuanchao Du

    Decoupling shared representations and task-specific knowledge for continual learning.

    arXiv PDF

  3. arXiv 2025

    TDR: Task-Decoupled Retrieval with Fine-Grained LLM Feedback for In-Context Learning

    Yifu Chen*, Bingchen Huang*, Zhiling Wang, Yuanchao Du, Junfeng Luo, Lei Shen, Zhineng Chen

    arXiv

  4. CVPR-W 2024

    NTIRE 2024 Quality Assessment of AI-Generated Content Challenge

    Xiaohong Liu, Xiongkuo Min, Guangtao Zhai, et al. (incl. Bingchen Huang)

    IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops

    arXiv

  5. ICCV 2023

    MRN: Multiplexed Routing Network for Incremental Multilingual Text Recognition

    Tianlun Zheng, Zhineng Chen, Bingchen Huang, Wei Zhang, Yu-Gang Jiang

    IEEE/CVF International Conference on Computer Vision

    arXiv PDF Code

  6. AAAI 2023

    Resolving Task Confusion in Dynamic Expansion Architectures for Class Incremental Learning

    Bingchen Huang, Zhineng Chen, Peng Zhou, Jiayin Chen, Zuxuan Wu

    Proceedings of the AAAI Conference on Artificial Intelligence

    arXiv AAAI Code

Ongoing Research

Agent self-evolution · Manuscript in preparation · 2026

GRSO: Group-Relative Skill Optimization for Frozen Agents

Optimizing agent skills with problem-level group sampling, contrastive reflection on successful and failed trajectories, and per-problem validation, while keeping model weights frozen.