CURRICULUM VITAE / 2026个人简历 / 2026
The work, at a glance.研究与经历,一目了然。
A concise public CV covering current manuscripts, research projects, and engineering. The work page and research notes tell the deeper story.这份公开版简历汇总在审论文、代表研究与工程经历。项目方法与思考详见代表工作和研究手记。
RESEARCH · ENGINEERING · AGENTS
Yidan Huang 黄一丹
Generative models, multimodal learning & agents生成模型、多模态学习与长程智能体
Researcher and engineer working across efficient reasoning, on-policy distillation, generative and multimodal models, and long-horizon agents. M.S. at CASIA / UCAS; LLM algorithm intern at JD Explore Academy.研究高效推理、在线蒸馏、生成与多模态模型,以及长程智能体。中国科学院自动化研究所 / 中国科学院大学硕士研究生,京东探索研究院大模型算法实习生。
JD Explore Academy京东探索研究院
JAN 2026 — PRESENTLLM Algorithm Intern · Natural Language Lab大模型算法实习生 · 自然语言实验室
- Work on long-context full-parameter SFT, preference optimization, and agent post-training workflows.参与长上下文全参数 SFT、偏好优化与智能体后训练流程。
- Connect trajectory provenance, supervision masks, packed sequences, checkpoint lineage, and agent evaluation.贯通轨迹来源、监督掩码、序列打包、权重血缘与智能体评测。
- Co-authored the JoyAI-LLM Flash report and supported its open release (48B total / ~2.7B active MoE).参与 JoyAI-LLM Flash 技术报告与模型开源发布(总参数 48B / 激活约 2.7B)。
UNDER REVIEW · ICLR 2027在审 · ICLR 2027ACTS · CATER · SP-OPD · CrystalBind · ClueFlow · SVI-Bench
UNDER REVIEW · AAAI 2027在审 · AAAI 2027TEAL · SearchAuditor
ACTS
Evidence-first synthesis links source-backed facts to the search actions needed to discover them.证据优先任务合成,连接有来源的事实与发现事实所需的检索动作。
CATER
Routes verified student rollouts to correctness RL or concise self-distillation by current competence.按学生当前能力,将验证过的采样分配给正确性强化学习或简洁自蒸馏。
SP-OPD
KL-projects teacher targets to respect verifier-accepted trajectories during on-policy distillation.以 KL 投影调整教师目标,让在线蒸馏尊重验证器认可的正确轨迹。
ClueFlow CO-FIRST AUTHOR共同第一作者
Persistent evidence and candidate state support provenance and turn-level credit in long search.以持久证据与候选状态支持长程搜索中的来源追踪和逐轮归因。
SVI-Bench CO-FIRST AUTHOR共同第一作者
A 75-item, nine-scenario streaming-video benchmark with five dimensions and synchronized InteractFlow traces.覆盖九类场景、75 个条目的流式视频基准,结合五维评分与同步交互轨迹。
TEAL
Typed evidence slots make fixed multimodal features readable to pretrained token backbones.用类型化证据槽,使固定多模态特征成为预训练 token 骨干可读取的输入。
SearchAuditor THIRD AUTHOR第三作者
Localizes critical errors in long search trajectories and grounds repair instructions in evidence.定位长程搜索轨迹中的关键错误,并以证据支撑修复建议。
CrystalBind UNDER REVIEW在审
Counterfactual prompt switches probe diffusion attribute binding; inference-time guidance targets ownership.用反事实提示词切换研究扩散模型属性绑定,并在推理时引导属性归属。
2026JoyAI-LLM Flash: Advancing Mid-Scale LLMs with Token Efficiency. Aichen Cai et al., including Yidan Huang. arXiv:2604.03044 ↗
2024Robust Detection of Cracked Eggs Using a Multi-Domain Training Method for Practical Egg Production. Yuxuan Cheng, Yidan Huang et al. Foods 13(15), 2313 ↗
2024A Study of a Photovoltaic Power Prediction Method Combining BiLSTM and GWO. Xiaokang Gui, Lewen Tan, Yidan Huang. ICCEA / IEEE ↗
CASIA / UCASM.S., Computer Application Technology · 2025—2028 (expected)计算机应用技术硕士 · 2025—2028(预计)
Huazhong Agricultural UniversityB.S., Information & Computing Science · 2021—2025
GPA 3.92/4.00 · Rank 1/102 · Two National Scholarships信息与计算科学学士 · 2021—2025
GPA 3.92/4.00 · 专业 1/102 · 两次国家奖学金
PyTorch · Diffusers · Transformers / PEFT · NeMo / FSDP2 · Megatron-Core · Slime · ms-swift · DeepSpeed · vLLM · SGLang · Kubernetes
Diffusion guidance · multimodal interfaces · SFT · DPO · RL / OPD · long-context agents · evaluation