关于我ABOUT
Benjamin Taoson 赖建铭
我关心 AI 如何从理解问题走向可靠行动。实践围绕两条方向展开：数字世界中的大模型与智能体，以及物理世界中的机器人学习。I am interested in how AI moves from understanding a problem to acting reliably. My work follows two directions: language models and agents in the digital world, and robot learning in the physical world.
这个网站收录项目的完整过程：问题怎么定义，为什么这样设计，做出了什么，以及实验结果说明了什么。This site documents the full story behind each project: how I frame the problem, why I choose a design, what I build, and what the experiments show.
两个方向，同一个问题Two directions, a shared question
一个 AI 系统怎样利用证据做出判断，并在判断错误时修正行动？这连接了我在数字 AI 与物理 AI 中的兴趣。How does an AI system use evidence to make a decision, and correct its actions when that decision is wrong? That question connects my interests in digital and physical AI.
Digital AI
从大模型到智能体系统，关注任务理解、工具使用、模型后训练与可靠性。项目涉及经营分析、多模态内容生产、领域微调和奖励建模。Language models and agent systems, with a focus on task understanding, tool use, post-training, and reliability. Projects cover business analysis, multimodal content, domain fine-tuning, and reward modeling.
数字 AI 项目Digital AI projects →Physical AI
关注机器人学习、VLA、世界模型与闭环决策。以 SpatialMind 的二维仿真为起点，探索空间记忆、主动感知，以及环境变化后的纠错与恢复。Robot learning, VLA policies, world models, and closed-loop decisions. SpatialMind's 2D simulation is a starting point for exploring spatial memory, active perception, and recovery when the environment changes.
SpatialMind →我怎样做项目How I approach a project
我习惯从具体任务出发，把问题拆成可以实现和验证的步骤。做完一个系统，还要回头检查：结果从哪里来，换一个条件是否成立，失败又说明了什么。项目页面记录这些过程，也保留做出技术选择时的判断。I start with a concrete task and break it into steps I can implement and test. After building a system, I ask where the result comes from, whether it holds under different conditions, and what failures reveal. The case studies document that process and the reasoning behind the design decisions.
把设计放回具体任务Let the task shape the design
在 AI BA 中，模型负责理解问题与安排调查，程序负责受控查询和确定性计算。这样的分工来自一个具体要求：分析结论需要能追溯到一致的指标定义和数据。查看 AI BA 的架构与取舍 →In AI BA, the model interprets questions and plans investigations, while code handles governed queries and deterministic calculations. This division follows a concrete requirement: conclusions must trace back to consistent metric definitions and data. Explore AI BA's architecture and decisions →
继续追问结果为什么成立Ask why a result holds
Reward Modeling Lab 得到较高的偏好判断准确率后，我继续检查长度捷径：一个总选更长回答的规则得分更高。长度匹配、反长度挑战和整体排序，帮助我更准确地解释模型学到了什么。查看实验与反思 →After a high preference score in Reward Modeling Lab, I checked for length shortcuts: an always-pick-the-longer-answer rule scored even higher. Length-matched, reversed-length, and ranking evaluations helped clarify what the model had learned. Read the experiments and reflections →
持续探索的问题Questions I keep exploring
评测得分接近的模型，面对关键证据变化时，会作出相似的判断吗？如果机器人记得的位置已经过时，它应该相信记忆，还是重新观察？我通过 RewardLens 与 SpatialMind 中的实验，逐步探索这些问题。Do models with similar evaluation scores behave similarly when key evidence changes? When a robot's remembered location becomes outdated, should it trust its memory or look again? Experiments in RewardLens and SpatialMind help me explore these questions.
写作与交流Writing and conversation
项目之外，我把技术文章、源码阅读与学习记录整理在独立的知识星球 ↗。欢迎带着一个工程问题、研究观察、合作想法或授课邀请来交流。Beyond the projects, my technical writing, source-code reading, and learning notes live in a separate knowledge hub ↗. I welcome engineering questions, research observations, collaboration ideas, and teaching invitations.
联系我Get in touch