About Me

Hi! I am Yi Hong (洪逸), a M.S. student in Software Engineering at the School of Computer Science and Engineering, Southeast University (东南大学). My research centers on Multi-Agent Reinforcement Learning and Knowledge Conflict Detection.

I am currently a agent development intern on the AI Platform (Agent Evaluation) team at ByteDance, where I build automated benchmark-production pipelines and LLM-driven trajectory-analysis tooling for evaluating coding agents. I am broadly interested in agent engineering, agentic RL, and retrieval-augmented systems.

News

  • 2026.10 Awarded the National Scholarship (国家奖学金).
  • 2026.08 Our paper accepted at NeurIPS 2026 (CCF-A).
  • 2026.05 Started as a agent development intern at ByteDance — AI Platform, Agent Evaluation.
  • 2025.10 Awarded the National Scholarship (国家奖学金).
  • 2025.09 Our paper accepted at IEEE TKDE (CCF-A).
  • 2024.09 Started my M.S. at Southeast University.

Education

Southeast University

M.S. in Software Engineering · School of Computer Science and Engineering

Admitted by recommendation

Southeast University

B.Eng. in Software Engineering · School of Software

Experience

ByteDance AI Platform (Agent Evaluation)

Agent Development Intern

Coding AgentAgent Evaluation

  • Coding Agent Benchmark construction: Built an automated benchmark-production pipeline (task generation → environment construction → F2P/P2P testing → online trial runs) on top of internal CLI tools and code repositories, producing high-quality, discriminative benchmarks free of data leakage — inspired by the SWE-Bench methodology — to drive model & product iteration for the TRAE team.
  • Automated analysis & report tooling: Designed an LLM-based automatic labeling workflow that identifies key behaviors in SOTA-model trajectories against preset rubrics, quantifies how weaker models deviate, and feeds behavior-grounded context to a report-generation agent.
  • Anti-cheating evaluation pipeline: Designed a dual-container (controller + executor) execution link that isolates the solving sandbox from evaluation artifacts, preventing agent cheating across multi-round evaluations.

Whale Cloud Technology (浩鲸科技) LLM Innovation Department

Algorithm Intern

LangChainGraphRAG

  • Built a usable knowledge graph from private-domain data to power structured retrieval for an upstream QA assistant; designed and iteratively optimized the ontology layer, reducing entity / relation missing-rates by 13.84% and 33.47%.
  • Preprocessed documents with the internal DocChain parser to build a document tree, enabling targeted sub-question decomposition and improving golden-chunk recall by 22.56%.

Publications

Effective Knowledge Conflict Detection via Joint Agentic Optimization

Yi Hong, Wenchao Bai, Jiaqi Jiang, Siyuan Huang and Jiahui Jin

Advances in Neural Information Processing Systems (NeurIPS), 2026 · CCF-A · First author

A dual-agent task-decomposition architecture that exploits the sparsity of knowledge conflicts to reduce detection complexity from O(m²) to O(m + Nk²), with <15% extra token cost and <8s latency. Jointly trained knowledge-grouping and conflict-detection agents with GRPO, achieving SOTA accuracy–efficiency trade-off across multiple datasets.

Framework overview of the knowledge-conflict detection method

An Event-Centric Framework for Predicting Crime Hotspots With Flexible Time Intervals

Jiahui Jin, Yi Hong, Guandong Xu, Jinghui Zhang, Jun Tang and Hancheng Wang.

IEEE Transactions on Knowledge and Data Engineering (TKDE), 2025 · CCF-A · Student first author

FlexiCrime, an event-centric framework for predicting urban crime hotspots over arbitrary, user-specified time intervals — instead of fixed time granularities. It combines a continuous-time attention network that learns general crime-context patterns with a type-aware spatiotemporal point process that captures crime-evolving risk per type, time and location. Experiments on real-world data from two cities across twelve crime types show it outperforms baselines on flexible-interval prediction.

Framework overview of FlexiCrime

Personal Projects

Agent History Viewer

Coding AgentCodexClaude CodeNode.js

  • A cross-platform local viewer that unifies scattered Codex / Claude Code conversation logs (across directories and JSONL formats) into a single searchable, traceable work archive — with a time-ordered timeline, cross-source / workspace / favorite filters, and both full-text and in-session search.
  • Provides a usage dashboard (token trends, activity heatmap, model & tool-call breakdown), paired tool-call/result rendering with sub-agent highlighting, Markdown / HTML export, resumable `.agenthistory.zip` session packages, and registration as a background service on macOS / Linux / Windows.

Daily Intelligent Assistant

LLM LoopAgent MemoryOpenClaw

  • A multi-channel assistant built on an LLM loop, unifying Feishu / QQ platforms via a message bus for automated task management and cross-platform message handling.
  • Integrated Shell execution and Web search tools with multi-turn tool chaining, plus a two-layer memory mechanism (long-term memory + history log) to improve contextual accuracy.

PyG Remote Backend based on GraphScope

Graph LearningAI InfraOSPP 2024

  • Provided distributed graph storage & sampling for PyG on top of GraphScope, implementing the FeatureStore and GraphStore data-access interfaces on the GraphScope client — Open Source Promotion Plan (开源之夏) 2024.

Honors & Awards

  • 2026.10 National Scholarship (国家奖学金)
  • 2025.10 National Scholarship (国家奖学金)
  • 2025.10 First-Class Graduate Academic Scholarship
  • 2025.10 Second Prize, Huawei Track — China Graduate AI Innovation Competition
  • 2024.10 Second-Class Graduate Academic Scholarship
  • 2023.11 National Third Prize — "Challenge Cup" (挑战杯 揭榜挂帅)
  • 2023.10 SEU Zhishan Scholar Scholarship (至善学子奖学金)

Skills

Languages: Python (primary); C++.

Workflow & Tools: Linux, VS Code, proficient with AI coding tools (Claude Code, etc.)

Contact