记录谢bro的研究随想和见闻。https://xjiajiahao.github.io/zh-CNhttps://xjiajiahao.github.io/articles/agent-architecture-tools-memory-and-reliable-execution/https://xjiajiahao.github.io/articles/agent-architecture-tools-memory-and-reliable-execution/从任务执行闭环出发，梳理 Harness 的执行控制、上下文与记忆、工具与协议、安全隔离、验证和恢复。Thu, 15 Oct 2026 00:00:00 GMTAgentHarness工具调用记忆https://xjiajiahao.github.io/articles/llm-fine-tuning-and-preference-optimization/https://xjiajiahao.github.io/articles/llm-fine-tuning-and-preference-optimization/比较 SFT、RLHF、PPO、DPO 与 GRPO 的数据、目标和工程边界，并建立是否应该微调的决策路径。Thu, 08 Oct 2026 00:00:00 GMT微调RLHFDPOGRPOhttps://xjiajiahao.github.io/articles/math-and-machine-learning-cheat-sheets/https://xjiajiahao.github.io/articles/math-and-machine-learning-cheat-sheets/汇总优化不等式、矩阵分析、概率论、Metropolis–Hastings 与最优传输中的常用定义、公式和推导。Sun, 04 Oct 2026 00:00:00 GMT数学机器学习概率论优化https://xjiajiahao.github.io/articles/rag-retrieval-systems/https://xjiajiahao.github.io/articles/rag-retrieval-systems/从企业知识库场景出发，梳理解析、索引、混合召回、融合、重排、引用与评测构成的完整证据链。Thu, 17 Sep 2026 00:00:00 GMTRAG检索向量数据库Agenthttps://xjiajiahao.github.io/articles/decoder-only-transformer-dataflow/https://xjiajiahao.github.io/articles/decoder-only-transformer-dataflow/沿着一次 token 生成，拆解 Decoder-only Transformer 的前向计算、训练推理、长上下文机制与 Agent 能力形成路径。Sun, 13 Sep 2026 00:00:00 GMTTransformerLLM长上下文