MEIHAO FAN

Curriculum Vitae

Download CV (PDF)

Research Summary

Ph.D. student at Renmin University of China researching scalable data synthesis for agent training. I develop methods for task generation, interaction-trajectory synthesis, and execution-based data verification, and use the resulting data for curriculum learning and multi-turn reinforcement learning. My work connects data construction to measurable improvements in agent capability and generalization. I am extending this research to embodied data curation, action-conditioned world models, and evaluation-driven data evolution.

I am seeking a Summer 2027 U.S. research internship in embodied AI data, world models, and multimodal agents.

Education

  • Renmin University of China, Beijing, China — Sept 2023 – Expected June 2028
    • Ph.D. in Computer Application Technology; advisors: Prof. Ju Fan and Prof. Xiaoyong Du.
    • National Scholarship (2025): the college’s only second-year Ph.D. recipient.
  • Chongqing Jiaotong University, Chongqing, China — Sept 2019 – June 2023
    • B.S. in Computer Science and Technology; GPA: 4.31/5.00.
    • National Scholarship (2023): the college’s only recipient; Mingde Scholarship Nomination Award (university top 20).

Research Experience

  • EvoPhys, Beijing, China — Research Intern, World Models — Sept 2026 – Present
    • Lead research on world model training, spanning interactive-memory evaluation, scalable multimodal data annotation, and evaluation-driven recursive self-improvement (RSI).
  • ByteDance, Beijing, China — Research Intern, LLM Agent Systems & Reinforcement Learning — March 2025 – June 2026
    • Led DeepPrep, covering synthetic training tasks, an executable data-preparation environment, and progressive SFT/RL; led research on agent harness self-evolution from execution trajectories.
  • Renmin University of China, Beijing, China — Research Assistant — Sept 2023 – March 2025
    • Led AutoPrep and BATCHER research on task-aware data preparation, multi-agent orchestration, demonstration selection, and cost-efficient LLM inference.

Research Projects

DeepAnalyze: Scalable Training-Trajectory Synthesis

ICML 2026 · Project Lead · Synthetic data, quality filtering, curriculum learning, agentic RL

  • Developed data-grounded trajectory synthesis for DeepAnalyze-8B, combining teacher-distilled reasoning and environment-generated interactions in the DataScience-Instruct-500K training corpus.
  • Built a questioner–solver–inspector pipeline: generate tasks and acceptance criteria from real data, execute multi-turn solutions, and filter trajectories using interaction checks and resulting environment changes.
  • Implemented a curriculum from single-skill SFT to multi-skill cold-start training and GRPO, using approximately 470K, 20K, and 15K samples, respectively.
  • Evaluated on 12 benchmarks; achieved 38.88% on DABStep and 61.7% on DS-1000, versus 15.34% and 54.8% for the single-ability training variant.

DeepPrep: Verifiable Task Synthesis for Agent Training

VLDB 2026 · Project Lead; First Author · Executable supervision, reversible corruption, multi-turn RL

  • Designed a synthesis method that converts SQL workloads into source tables, target specifications, and executable pipelines; adds reversible noise and verifies recovery to preserve task solvability.
  • Constructed 6,788 training tasks, with pipelines spanning 31 operator types and up to 28 steps; trained agents through operator-level SFT, verified reasoning trajectories, and GRPO with hybrid rewards.
  • Synthetic training data improved out-of-domain accuracy by 37.0 percentage points over Parrot training on Synth-Bird (Qwen3-8B, data ablation); tree-based execution supports backtracking and pipeline repair.

Interactive Memory & Recursive Self-Improvement (RSI) for World Models

Ongoing · Project Lead · World model evaluation, scalable data curation, training, self-improvement

  • Proposed an interactive-memory benchmark for action-conditioned world models, formalizing memory through Encoding, Maintain, Update, and Read and evaluating reactive response, memory persistence, and memory plasticity across temporal-retention and interference settings.
  • Built an automated Isaac Sim evaluation pipeline with programmatically generated cases and ground-truth state supervision; reproduced and evaluated nine representative world models, including Ctrl-World, WorldMem, iVideoGPT, HyDRA, and Oasis, under a unified evaluation protocol.
  • Built a fully automated multimodal annotation system for world-model training, covering atomic action segmentation, semantic and object-interaction labels, object-state changes, camera pose, and 3D hand motion; supports 100+ videos concurrently at approximately $40 API cost per raw video hour, with annotation quality validated through manual audits.
  • Developing a Recursive Self-Improvement pipeline linking data annotation, training-recipe construction, distributed training, evaluation, and recipe refinement; uses model weaknesses to drive hard-example mining, curriculum construction, active data selection, and data-mixture optimization.

CoDA-Bench & TACO: Scalable, Verifiable Task Construction

ICML / VLDB 2026 · Co-author · Task synthesis, execution verification, difficulty refinement

  • CoDA-Bench: Co-authored a benchmark of 1,009 tasks across 31 data communities, built from executable notebook solutions through task reconstruction, adversarial refinement, and human verification.
  • TACO: Co-authored a benchmark combining 1,500 expert-annotated and 13,000 synthetic NL–SQL pairs, using query-structure sampling, schema-constrained generation, and quality checks.

EvoCost & SkillAdam: Agent Self-Evolution

ICLR 2027 submissions · EvoCost: Lead, First Author · SkillAdam: Co-author

  • EvoCost: Developed harness evolution that mines avoidable interactions from trajectory dependency graphs and validates updates to instructions, skills, tools, and executor logic with a fixed backbone model.
  • SkillAdam: Co-authored skill optimization with persistent issue memory and adaptive edit budgets; reduced optimization tokens by 67.3% versus SkillOpt on DeepPlanning, with test accuracy improving from 21.7% to 28.3%.

AutoPrep & BATCHER: Task-Aware Data Preparation

VLDB 2025 / ICDE 2024 · Project Lead; First Author · Data transformation, demonstration selection, inference efficiency

  • AutoPrep: Developed Chain-of-Clauses planning and tool-augmented execution; improved NL2SQL accuracy by 12.22 / 13.23 points on WikiTQ / TabFact, averaged across evaluated backbones.
  • BATCHER: Developed coverage-based demonstration selection and token-aware batch allocation; batch prompting reduced API cost by 4–7× versus standard prompting across eight ER datasets.

Publications

  • Meihao Fan, Ju Fan, Yuxin Zhang, Shaolei Zhang, Xiaoyong Du, Jie Song, Peng Li, Fuxin Jiang, Tieying Zhang, Jianjun Chen. DeepPrep: An LLM-Powered Agentic System for Autonomous Data Preparation VLDB 2026. Paper

  • Meihao Fan, Ju Fan, Nan Tang, Lei Cao, Guoliang Li, Xiaoyong Du. AutoPrep: Natural Language Question-Aware Data Preparation with a Multi-Agent Framework VLDB 2025. Paper

  • Meihao Fan, Xiaoyue Han, Ju Fan, Chengliang Chai, Nan Tang, Guoliang Li, Xiaoyong Du. Cost-Effective In-Context Learning for Entity Resolution: A Design Space Exploration ICDE 2024. Paper

  • Shaolei Zhang, Ju Fan, Meihao Fan, Guoliang Li, Xiaoyong Du. DeepAnalyze: Agentic Large Language Models for Autonomous Data Science ICML 2026. Paper · Code

  • Yuxin Zhang, Meihao Fan, Ju Fan, Mingyang Yi, Yuyu Luo, Jian Tan, Guoliang Li. Reward-SQL: Boosting Text-to-SQL via Stepwise Execution-Aware Reasoning and Process-Supervised Rewards SIGMOD 2026. Paper

  • Yuxin Zhang, Ju Fan, Meihao Fan, Shaolei Zhang, Xiaoyong Du. CoDA-Bench: Can Code Agents Handle Data-Intensive Tasks? ICML 2026. Paper

  • Chao Deng, Ju Fan, Yuyu Luo, Qinliang Xue, Meihao Fan, Yuxin Zhang, Min Zhang, Xiaofeng Jia, Jing Zhang, Xiaoyong Du. TACO: A Benchmark for Open-Domain Text-to-SQL with Ambiguous and Cross-Database Queries VLDB 2026. Paper

Manuscripts

  • Meihao Fan, Shaolei Zhang, Ju Fan, Peng Li, Jie Song, Jianjun Chen. EvoCost: Harness Self-Evolution for Cost-Efficient LLM Agents Under Review at ICLR 2027.

  • Gaoyuan Li, Meihao Fan, Yizhe Liu, Shaolei Zhang, Ju Fan, Siyi Wang, Jiaheng Hou, Xudong Weng, Honghan Tian, Zang Li. SkillAdam: Stable and Efficient Skill Evolution for Agents Under Review at ICLR 2027. Preprint · Code

  • Yuxin Zhang, Ju Fan, Meihao Fan, Shaolei Zhang, … Multi-Agent World: Scaling Multi-Agent Orchestration Training via Recursive Environment Improvement Under Review at ICLR 2027.

Technical Skills & Research Interests

  • Data: Task and trajectory synthesis, execution-based verification, quality filtering, demonstration selection.
  • Training: Supervised fine-tuning, GRPO, multi-turn RL, curriculum learning, reward modeling.
  • Tools: Python, PyTorch, Transformers, vLLM.
  • Languages: Chinese (native), English (professional).
  • Research interests: Synthetic data for agents; embodied data; world models; multimodal agents; self-improving systems.