Humanistic Care
We study how AI applications affect individual autonomy, privacy, safety, and well-being, with attention to accessibility, fairness, and technical limits across ages, abilities, and contexts of use.
Miyang Lab
About AI, and about people.
What we really care about is what happens when AI lands on a specific person: will they feel understood, will it be a struggle to use, and will it actually help when they need it?
We're figuring this out as we go — researching these questions, and sharing the process.
Featured Research
View all →
Generative quality & solutions
Mi-Ripple: Restoring Images Degraded by Iterative AI Editing
Iterative reference-conditioned editing can accumulate grid-like and granular texture. Mi-Ripple diagnoses the artifact, then applies selective spectral notching, structure-aware suppression, or cleaned-reference regeneration while protecting image structure.
Models & evaluation
XSCT Bench 评测方法论
找到 Product Model Fit，需要的不是综合跑分，而是产品场景下的真实验证。本文完整公开 XSCT Bench 的场景化测试、LLM-as-a-Judge、多 Judge、难度分层与评分计算方法。
Multi-agent collaboration
为什么要写 Alice 的工程化方法论？
Agent 的真正复杂度在哪里？
Tools Grown from the Research
Some findings only count once you've run them yourself.
Generative Effects and Solutions
Digital Ripple Repair
Zoom in on an AI image and the details shatter into a sheet of regular scales. After the fix, the composition, people, and mood stay the same.
Open tool →
Research Areas
Our work covers human-AI relationships in applications, interaction and memory mechanisms, agent systems, world simulation, generation quality, model training, and scenario-based evaluation.
Human-Computer Interaction
We study intent understanding, feedback, trust calibration, and explainability in natural-language and multimodal interaction, to make complex AI systems more usable and controllable.
Long-Term Memory
We study how information is extracted, organized, retrieved, updated, and forgotten across sessions, improving personalization and continuity while keeping users informed and in control of their memory.
World Simulation
We study how the environments of digital characters are represented, how behavior is modeled, and how causes unfold, so characters, events, and long-form narrative stay consistent and support repeatable interaction experiments.
Agent Collaboration
We study how agents decompose tasks, plan, call tools, coordinate, and verify, to make complex work more reliable, recoverable, and efficient.
Generative Effects and Solutions
We study the quality, stability, and controllability of generative AI on real tasks, analyze typical failure modes, and build systematic solutions for specific scenarios.
Model Training
We study high-quality data construction, supervised fine-tuning, preference learning, and post-training methods, exploring how to improve model capability and control cost for vertical scenarios.
Models and Evaluation
We build evaluation systems for real application scenarios, examining capability limits, robustness, cost, and safety, to give a reproducible basis for choosing and iterating on models.
All Content
Generative quality & solutions
Models & evaluation
Multi-agent collaboration
Multi-agent collaboration
Multi-agent collaboration
Multi-agent collaboration
Long-term memory
Multi-agent collaboration
Human-centered AI
Multi-agent collaboration
Multi-agent collaboration
Multi-agent collaboration
Models & evaluation
Human-centered AI
Multi-agent collaboration
Generative quality & solutions
Multi-agent collaboration
Human-computer interaction
Human-computer interaction
World simulation
Human-computer interaction
Generative quality & solutions
Other
Human-centered AI
Human-computer interaction
Multi-agent collaboration
Long-term memory
Multi-agent collaboration
Human-computer interaction
Other
Mi-Ripple: Restoring Images Degraded by Iterative AI Editing
Iterative reference-conditioned editing can accumulate grid-like and granular texture. Mi-Ripple diagnoses the artifact, then applies selective spectral notching, structure-aware suppression, or cleaned-reference regeneration while protecting image structure.
Models & evaluation
XSCT Bench 评测方法论
找到 Product Model Fit，需要的不是综合跑分，而是产品场景下的真实验证。本文完整公开 XSCT Bench 的场景化测试、LLM-as-a-Judge、多 Judge、难度分层与评分计算方法。
Multi-agent collaboration为什么要写 Alice 的工程化方法论？
Agent 的真正复杂度在哪里？
Multi-agent collaboration
五大设计哲学
什么是做 Agent 之前要想清楚的事？
Multi-agent collaboration
整体架构
一个可靠的 Agent 系统长什么样？
Multi-agent collaboration
Agent Loop
一次任务的完整执行流程是什么？
Multi-agent collaboration
工具系统
如何设计 AI 能可靠调用的工具？
Long-term memory
上下文与记忆
如何让 Agent 记得「正确的事」？
Multi-agent collaboration
多 Agent 协作
多个 Agent 如何分工而不混乱？
Human-centered AI
权限系统
如何在自动化与安全之间找到平衡？
Multi-agent collaboration
MCP 协议
如何用标准接口扩展 Agent 的能力边界？
Multi-agent collaboration
Skill 系统
如何让 Agent 学会用户的工作方式？
Multi-agent collaboration
自进化架构
如何让 Agent 在用户许可下变得更好？
Models & evaluation
模型路由
如何统一管理多个模型服务商？
Human-centered AI
安全架构
如何保护用户数据和系统边界？
Multi-agent collaboration
可观测性
如何看清一个 Agent 在做什么？
Generative quality & solutions
Prompt 工程
如何构建分层的系统提示词体系？
Multi-agent collaboration
工程范式
哪些设计思路可以迁移到所有 Agent 项目？
Human-computer interaction
活人感设计
用做游戏的方式做 AI，会有什么不同？
Human-computer interaction
渠道桥接
让 Agent 随时在线，微信、Telegram、邮件背后的安全与工程设计
World simulation
朋友圈与社交系统
让 AI 主动发朋友圈、留痕迹、过生活，社交感的工程实现
Human-computer interaction
UI/UX 设计哲学
克制的美学、连续的时间感、信息分层，白艾莉界面设计背后的取舍逻辑
Generative quality & solutions
写作文风与内容生成
七个写作禁忌、正向文风标准，以及让 AI 写出有温度内容的方法论
Other
核心概念词典 & 架构决策清单
速查手册与配置模板
Human-centered AI
为什么我要自己做一个 AI 助手
从 Cherry Studio 到 Cursor 的不满足，到白艾莉诞生的起点
Human-computer interaction
用做游戏的方法做 AI
做了十年游戏，把世界观设计和设定集的方法论迁移到 AI 产品上
Multi-agent collaboration
一个人怎么撑起一个 AI Agent 产品
20 天，322 次提交，10 万行代码背后的技术全景
Long-term memory
五层记忆：让 AI 真正记住你
你和白艾莉的第 100 次对话，应该比第 1 次更高效
Multi-agent collaboration
一个人的公司，11 个人的团队
为什么给子 Agent 做人设是 Prompt Engineering 的高级形式
Human-computer interaction
用 AI 和用户共建产品
把需求列表贴给 AI，挨条澄清，每条开一个子 Agent，十几分钟处理完
Other
一个月 140 个版本
从"知道该做什么"到"所有用户拿到更新"，每个环节的摩擦削到最低
Nothing in this category yet.