陈宇卓Yuzhuo Chen

QUESTIONS & INTERESTS · CURIOSITY-DRIVEN EXPLORATION

我好奇的问题与探索Questions & Explorations

这里记录我持续好奇的技术兴趣与研究问题：已有工作走到哪里、不同路线怎样权衡，以及我更认可什么。兴趣综述不意味着原创声明，研究设想也不是已验证结论。Technical interests and research questions: prior work, trade-offs, and current perspectives. An interest survey is not a novelty claim, and proposed ideas are not established findings.

INTEREST 01 兴趣专题Interest survey

如何让模型学会自适应推理How Can Models Learn Adaptive Reasoning?

为什么自适应推理影响 Agent 的固定预算有效产出与机器人在线时延？梳理 SFT、成本敏感 RL、步骤级信用分配、反应—规划切换的最新研究；区分真正学会思考时机与仅减少 CoT，并讨论有限推理机会能否促进参数化直觉的待验证假设。A critical survey of SFT, cost-aware RL, reasoning credit assignment, adaptive planning and robot latency. Tests the distinction between shorter CoT and genuinely learned, transferable parametric intuition.

技术路线综述 · 个人判断Technical survey · personal perspective 阅读技术路线、文献与思考 →Read survey and perspectives →
INTEREST 02 兴趣专题Interest survey

一个 Agent 是如何组织推理、工具调用与跨轮状态的？How Do Agents Organize Reasoning, Tool Calls, and Cross-Turn State?

从模型 Token、API 结构化输出到 Harness 的工具执行和状态回传，比较 Claude Messages、OpenAI Responses/Codex、Qwen/vLLM、OpenHands、LangGraph、smolagents 与 OpenCode；解释 Thinking、Tool Call、执行反馈和跨轮历史分别由谁负责。An end-to-end comparison of how Claude, OpenAI Responses/Codex, Qwen/vLLM, OpenHands, LangGraph, smolagents and OpenCode organize reasoning, tool use, execution feedback and conversation state.

技术路线综述 · 个人判断Technical survey · personal perspective 阅读技术路线、文献与思考 →Read survey and perspectives →
QUESTION 03 持续探索Exploring

目前的 Tokenizer，是否天然具备数值连续性？Do Today's Tokenizers Inherently Preserve Numerical Continuity?

普通语言 Tokenizer 不保证相近数值有相近表示；VLM 的连续视觉特征也不等于坐标输出具备连续性，而离散 Action Codebook 同样可能扭曲物理邻近关系。对照 xVal、BitTokens、Pix2Seq、FAST、X-Tokenizer、ActionPiece、ProAct 和连续动作头，提出几种可验证的改进路径。Text tokenization does not inherently preserve numerical distance. Continuous visual features do not guarantee metric output continuity, and discrete VLA action codes may distort physical neighborhoods. Review numerical encodings, action codecs, continuous heads, and testable research directions.

研究疑问 · 不是结论Open question · not a finding 阅读问题、文献与实验构想 →Read evidence and experiments →
QUESTION 04 持续探索Exploring

不同机器人数据集的动作采样频率，是否会引入数据集偏置？Do heterogeneous robot action rates induce dataset-specific shortcuts?

综合核查 30 组公开机器人数据源与采集管线、RDT-1B 的 46 个预训练数据源，以及 GR00T N1 的预训练 FPS：各来源配置覆盖 2–50 Hz，但与原始轨迹实测 Hz 不等价。异构混训确实存在，模型是否因此学到数据集捷径仍待控制实验。Reviewing 30 public robot data sources and pipelines, RDT-1B's 46 active pretraining sources and GR00T N1's per-source FPS reveals 2–50 Hz configurations. These are not necessarily original measured action rates; learned dataset shortcuts remain unproven.

研究疑问 · 不是结论Open question · not a finding 阅读问题、文献与实验构想 →Read evidence and experiments →

完整引用和研究过程只维护在研究知识库的对应 Markdown 中，避免重复更新。Detailed evidence is maintained in one canonical Research Library note.