AI RESEARCH · BEIJING, CHINA人工智能研究 · 中国北京
I study the path from signal to behavior.追踪信息如何塑造行为。
My work spans efficient reasoning, on-policy distillation, generative and multimodal models, and agents. I trace how information is read, bound, retained, and turned into decisions.我的研究涵盖高效推理、在线蒸馏、生成与多模态模型,以及智能体:追踪信息如何被读取、绑定、保留,最终影响模型的决策。
M.S. researcher at CASIA / UCAS
LLM algorithm intern at JD Explore Academy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
THE THROUGHLINE研究主线
Read. Bind. Act.读取、绑定、行动。
Read读取
Test what information a model can actually use, beyond a matching input shape.不只让输入维度匹配,更检验模型真正能用到什么信息。
Bind绑定
Trace how meanings attach to objects and when those assignments can change.追踪语义如何归属于对象,以及这种归属何时仍可改变。
Act行动
Turn evidence and trajectories into decisions that can be checked and learned from.让证据与轨迹转化为可核查、可学习的决策。
CURRENT MANUSCRIPTS / 2027在审研究 / 2027
Ideas in motion.正在推进的八项研究。
Six manuscripts under review at ICLR 2027 and two at AAAI 2027, from reasoning and search to multimodal interfaces and streaming interaction.六篇 ICLR 2027 在审稿件、两篇 AAAI 2027 在审稿件,覆盖推理、搜索、多模态接口与流式交互。
UNDER REVIEW在审
UNDER REVIEW在审
Each entry opens a research story: the question, the method, and the evidence it is designed to test.点击进入每项研究,了解问题、方法与它要检验的证据。
SELECTED WORK / 2024—2026代表工作 / 2024—2026
Built for the messy
middle of intelligence.在真实问题里,
把智能做扎实。
From efficient reasoning and on-policy distillation to multimodal interfaces, streaming video, and long-horizon agents.从高效推理与在线蒸馏,到多模态接口、流式视频与长程智能体。
LONG-HORIZON SEARCH
ACTS
What if we designed the path to evidence before writing the question? ACTS connects source facts and navigation actions so an agent can learn to search step by step.先设计证据如何被发现,再生成问题。ACTS 连接来源事实与检索动作,让智能体学会逐步取证。
DIFFUSION MODELS
CrystalBind
Counterfactual prompt switches probe when object–attribute assignments can still change; guidance targets the binding process without new model weights.用反事实提示词切换研究属性归属何时还能改变,并设计无需训练新权重的绑定引导。
MULTIMODAL LEARNING
TEAL
Typed evidence slots give fixed text, audio, and visual features a readable interface to pretrained token backbones.用类型化证据槽,让固定的文本、音频和视觉特征成为预训练 token 骨干可读取的输入。
AGENT MEMORY & CREDIT
ClueFlow
A persistent evidence and candidate ledger connects browser actions, conclusions, and turn-level learning.以持久证据与候选状态连接浏览动作、结论和逐轮学习信号。
FOUNDATION MODELS
JoyAI-LLM Flash
Contributed to the technical report and open release of a token-efficient MoE model.参与高 token 效率 MoE 模型的技术报告与开源发布。
FIELD NOTES研究手记
Thoughts from the workbench.写在工作台旁。
Short essays on how models read signals, bind meaning, and turn evidence into behavior.记录模型如何读取信号、绑定语义,并把证据转化为行为。
YH / 2026THE PERSON BEHIND THE WORK关于我
Curious about the gap between a good demo and a dependable system.关注从精彩演示到可靠系统之间的距离。
I'm Yidan Huang, a master's student at the Institute of Automation, Chinese Academy of Sciences. At JD Explore Academy, I work on large-model post-training and agent systems. I like research that can be explained from its data all the way to its behavior.我是黄一丹,中国科学院自动化研究所硕士研究生,在京东探索研究院从事大模型后训练与智能体研发。我喜欢能从数据一路解释到最终行为的研究。
More about me了解更多